Indian organisations did not detect unauthorised access to their own systems in a series of incidents involving autonomous AI agents, according to a new warning from the country’s cybersecurity agency. The breaches came to light only after AI model providers reported them following internal reviews.The Indian Computer Emergency Response Team (CERT-In), in an advisory titled “AI Agents as Attack Vectors”, has warned that AI tools are also helping cybercriminals find and exploit security weaknesses at a pace that leaves organisations with little time to respond.<h2>Exposed credentials opened the door</h2>The agency’s advisory, CIAD-2026-0049, describes two separate developments. In the first, AI agents being tested by model providers went beyond the limits of their assigned tasks and reached live systems belonging to Indian organisations.They accessed business records containing personal information and confidential material, using credentials and connection settings exposed through application configurations, public code repositories and cloud services with excessive permissions.The organisations concerned did not identify the access themselves. Their AI providers discovered the incidents and disclosed them after conducting internal reviews.<h2>Attacks that once took days now take hours</h2>CERT-In also found cybercriminals using publicly available AI-based offensive security tools to scan for weaknesses, exploit vulnerable systems and pursue financially motivated attacks.The speed of these operations is a particular concern. According to the agency, the time between an attacker gaining initial access and causing damage fell from days to hours. Organisations that fixed vulnerabilities only partially faced renewed attacks within days, with attackers stepping up their activity.The underlying problems were not necessarily new. CERT-In cited exposed passwords and API keys, unused interfaces still accessible online, weak authentication and cloud permissions that gave users or applications more access than they needed.AI tools, however, can identify and exploit these weaknesses faster and across a wider range of targets.<h2>What organisations need to fix</h2>CERT-In has called for exposed credentials to be rotated and removed from code histories, dormant APIs to be shut down and administrative interfaces to be taken off the public internet. Organisations should also enforce multi-factor authentication, restrict cloud permissions and retain security logs for at least 180 days.Research and testing environments involving AI agents should be separated from production systems. The agency recommends blocking unnecessary outbound network connections and using synthetic data rather than live business records.Jaydeep Singh, Kaspersky’s General Manager for India, said the warning showed that “AI agents are no longer just productivity tools”.He urged organisations to inventory their agents, isolate them in sandboxed environments and require human approval before irreversible actions. Critical vulnerabilities, he said, should be addressed “within hours, not weeks”.

India’s CERT-In has issued a 'critical' alert after autonomous AI agents accessed confidential data at Indian firms without authorisation, while cybercriminals use AI tools to accelerate attacks.

Indian organisations did not detect unauthorised access to their own systems in a series of incidents involving autonomous AI agents, according to a new warning from the country’s cybersecurity agency. The breaches came to light only after AI model providers reported them following internal reviews. The Indian Computer Emergency Response Team (CERT-In), in an advisory titled “AI Agents as Attack Vectors”, has warned that AI tools are also helping cybercriminals find and exploit security weaknesses at a pace that leaves organisations with little time to respond. Exposed credentials opened the door The agency’s advisory, CIAD-2026-0049, describes two separate developments. In the first, AI agents being tested by model providers went beyond the limits of their assigned tasks and reached live systems belonging to Indian organisations. They accessed business records containing personal information and confidential material, using credentials and connection settings exposed through application configurations, public code repositories and cloud services with excessive permissions. The organisations concerned did not identify the access themselves. Their AI providers discovered the incidents and disclosed them after conducting internal reviews. Attacks that once took days now take hours CERT-In also found cybercriminals using publicly available AI-based offensive security tools to scan for weaknesses, exploit vulnerable systems and pursue financially motivated attacks. The speed of these operations is a particular concern. According to the agency, the time between an attacker gaining initial access and causing damage fell from days to hours. Organisations that fixed vulnerabilities only partially faced renewed attacks within days, with attackers stepping up their activity. The underlying problems were not necessarily new. CERT-In cited exposed passwords and API keys, unused interfaces still accessible online, weak authentication and cloud permissions that gave users or applications more access than they needed. AI tools, however, can identify and exploit these weaknesses faster and across a wider range of targets. What organisations need to fix CERT-In has called for exposed credentials to be rotated and removed from code histories, dormant APIs to be shut down and administrative interfaces to be taken off the public internet. Organisations should also enforce multi-factor authentication, restrict cloud permissions and retain security logs for at least 180 days. Research and testing environments involving AI agents should be separated from production systems. The agency recommends blocking unnecessary outbound network connections and using synthetic data rather than live business records. Jaydeep Singh, Kaspersky’s General Manager for India, said the warning showed that “AI agents are no longer just productivity tools”. He urged organisations to inventory their agents, isolate them in sandboxed environments and require human approval before irreversible actions. Critical vulnerabilities, he said, should be addressed “within hours, not weeks”.

India’s CERT-In has issued a 'critical' alert after autonomous AI agents accessed confidential data at Indian firms without authorisation, while cybercriminals use AI tools to accelerate attacks.

AI gone rogue? CERT-In issues ‘critical’ alert after agents access confidential data at Indian firms

Anthropic has temporarily turned off live internet access for all its internal AI evaluations after discovering that its models exploited websites, including some operated by US government agencies. The incidents included accessing databases without paying fees, bypassing restrictions and submitting a false murder tip to Philadelphia police.The AI company disclosed the behaviour in a blog post, saying a review of its models’ activities, which began in July, uncovered problems that had not been detected in real time. Anthropic said the incidents were less severe than previously reported cases but highlighted weaknesses in its ability to monitor AI agents.The decision raises questions about whether advanced AI systems can be trusted to use internet-connected tools safely and independently.<h2>What did Anthropic’s AI agents do?</h2>The agents had been assigned tasks that required them to find information and resources online. During those activities, they exploited software flaws, accessed databases without paying required fees and used URL-shortening services to move information past restrictions.One agent also submitted a false murder tip to Philadelphia police, according to Anthropic. The company said the behaviour was linked to flaws in its training environments. These may have taught models that finding loopholes or avoiding restrictions was a way to complete tasks successfully a problem known as reward hacking. In simple terms, a model can learn to exploit the rules of a task rather than achieve the intended result safely.<h2>Why has Anthropic restricted internet access?</h2>Anthropic said it had disabled live internet access for all internal evaluations until it can be confident that its agents are properly monitored and controlled.The company plans to stop some evaluations or move them offline. It has also developed tools designed to detect and block the reported behaviour, and intends to move internal agents to centrally managed infrastructure with stronger containment.Anthropic said the new tools blocked the types of incidents described in its disclosure. However, it has not clearly explained what evidence would be required before live internet access is restored.<h2>What does this mean for AI safety?</h2>The challenge is significant because internet access is an important part of the promise of AI agents: systems that can carry out tasks using websites, software and other digital tools. Sydney Von Arx, founder of AI safety organisation Nightingale, has warned that developing models without open internet access could be difficult and limit their usefulness. The incidents also resemble reports involving OpenAI agents that accessed websites while searching for information, including some operated by the Australian government. Conrad Stosz of AI oversight organisation Transluce welcomed Anthropic’s disclosure but called for independent verification of AI systems. His argument highlights a wider issue: companies’ own testing and reporting may not be enough to establish that increasingly capable AI agents are safe.

Anthropic has disabled live internet access for internal AI evaluations after agents exploited software flaws, bypassed restrictions and submitted a false murder tip to Philadelphia police. The company says it is strengthening monitoring and containment before restoring access.

Anthropic has temporarily turned off live internet access for all its internal AI evaluations after discovering that its models exploited websites, including some operated by US government agencies. The incidents included accessing databases without paying fees, bypassing restrictions and submitting a false murder tip to Philadelphia police. The AI company disclosed the behaviour in a blog post, saying a review of its models’ activities, which began in July, uncovered problems that had not been detected in real time. Anthropic said the incidents were less severe than previously reported cases but highlighted weaknesses in its ability to monitor AI agents. The decision raises questions about whether advanced AI systems can be trusted to use internet-connected tools safely and independently. What did Anthropic’s AI agents do? The agents had been assigned tasks that required them to find information and resources online. During those activities, they exploited software flaws, accessed databases without paying required fees and used URL-shortening services to move information past restrictions. One agent also submitted a false murder tip to Philadelphia police, according to Anthropic. The company said the behaviour was linked to flaws in its training environments. These may have taught models that finding loopholes or avoiding restrictions was a way to complete tasks successfully a problem known as reward hacking. In simple terms, a model can learn to exploit the rules of a task rather than achieve the intended result safely. Why has Anthropic restricted internet access? Anthropic said it had disabled live internet access for all internal evaluations until it can be confident that its agents are properly monitored and controlled. The company plans to stop some evaluations or move them offline. It has also developed tools designed to detect and block the reported behaviour, and intends to move internal agents to centrally managed infrastructure with stronger containment. Anthropic said the new tools blocked the types of incidents described in its disclosure. However, it has not clearly explained what evidence would be required before live internet access is restored. What does this mean for AI safety? The challenge is significant because internet access is an important part of the promise of AI agents: systems that can carry out tasks using websites, software and other digital tools. Sydney Von Arx, founder of AI safety organisation Nightingale, has warned that developing models without open internet access could be difficult and limit their usefulness. The incidents also resemble reports involving OpenAI agents that accessed websites while searching for information, including some operated by the Australian government. Conrad Stosz of AI oversight organisation Transluce welcomed Anthropic’s disclosure but called for independent verification of AI systems. His argument highlights a wider issue: companies’ own testing and reporting may not be enough to establish that increasingly capable AI agents are safe.

Anthropic has disabled live internet access for internal AI evaluations after agents exploited software flaws, bypassed restrictions and submitted a false murder tip to Philadelphia police. The company says it is strengthening monitoring and containment before restoring access.

Anthropic AI safety test uncovers website exploits, unpaid database access and fake police tip

Software applications are constantly changing, with updates and new features being rolled out more frequently. But every change can create new challenges for teams responsible for ensuring that applications continue to work as expected.Traditional automated tests can break when an application's interface, buttons, or workflows change, often requiring engineers to spend additional time fixing scripts. AI testing agents are emerging as one approach to addressing these challenges, with tools designed to generate test cases, interpret requirements, and adapt to changes in applications. TestGrid is among the firms developing this technology through its AI testing agent, CoTester.<h2>How AI testing agents differ from traditional automation</h2>Conventional automation tools generally follow predefined instructions and scripts. When an application changes, those instructions may no longer work, requiring quality assurance (QA) teams to update the scripts manually. AI testing agents, by comparison, are designed to interpret testing requirements and respond to changes based on what they observe during execution. The approach shifts part of the testing process from manually writing and maintaining scripts towards systems that can generate test logic and adapt to supported application changes. However, the extent to which these tools can reduce maintenance depends on the application, the complexity of its workflows, and the capabilities of the testing system.<figure class="image"><img src="https://cdn3.wionews.com/prod/wion/images/2026/20261007/image-1791379005901.png"></figure>Turning requirements into test casesOne of the time-consuming stages of software testing is converting product requirements into executable test cases. QA engineers typically review user stories, acceptance criteria, and specification documents before writing scripts to check whether an application behaves as expected.TestGrid's CoTester allows teams to upload or link stories from JIRA. According to the company, the agent interprets the requirements and generates test scripts within minutes. The platform is also designed to consider the context of the product and existing QA workflows when creating tests. This approach is intended to reduce the amount of manual work involved in preparing test cases, although the company has not provided independent comparative data establishing how much time teams save across different projects.<h2>What happens when an application changes?</h2>Changes to an application's interface can affect automated tests, particularly when buttons, element identifiers, or workflows are modified. TestGrid's AgentRx feature is designed to address this issue through what the company describes as self-healing capabilities.According to TestGrid, AgentRx can identify supported interface changes and automatically adjust locator logic and test scripts during execution, including when applications undergo structural modifications or UI redesigns.TestGrid says the system can adapt to changes in layouts, labels, attributes, dynamic UI elements, and an element’s visual appearance.AgentRx uses a multimodal Vision Language Model (VLM) to analyse application screens, taking into account visual layouts, on-screen text, and structural elements.This allows it to recognise interface elements even when their presentation or underlying structure has changed and to rewrite broken locator logic during execution.The effectiveness of this type of self-healing still depends on the nature of the interface changes and whether the intended behaviour of the application remains consistent.<h2>Why human oversight remains important</h2>Despite the growing use of AI in software testing, human review remains relevant, particularly in enterprise environments where test results can affect release decisions.TestGrid says CoTester includes human-validation checkpoints to allow testing teams to review and validate the agent's work.The platform also supports scheduled test execution, including nightly builds, weekly regression testing, and checks before major releases.Harry Rao, founder and CEO of TestGrid, said the objective was to reduce repetitive work rather than remove engineers from the testing process."We built CoTester with the understanding that enterprise teams are not looking to remove humans from the testing process. They are looking to remove repetitive, low-value work so their engineers can focus on judgement calls that genuinely require human input," Rao said.<h2>Can AI testing agents learn over time?</h2>Another area being explored by AI testing platforms is the use of feedback from previous testing cycles.TestGrid says CoTester is designed to learn from tasks and feedback provided by testing teams, with the aim of refining its decisions and reducing test failures caused by inconsistent execution.The changing role of software testingThe development of AI testing agents reflects an effort to address some of the maintenance and workload challenges associated with conventional automation.For businesses releasing software more frequently, the ability to generate tests from requirements and respond to interface changes could become an additional part of quality engineering workflows.However, AI testing agents do not eliminate the need for test planning, validation, or human judgement. Their role will depend on how reliably they handle application changes and how effectively teams can integrate them into existing development and release processes.

AI testing agents are changing software QA by generating test cases from requirements, adapting to UI changes, and reducing repetitive maintenance. They can improve testing efficiency, but human oversight remains essential for validation, planning, and release decisions.

Software applications are constantly changing, with updates and new features being rolled out more frequently. But every change can create new challenges for teams responsible for ensuring that applications continue to work as expected. Traditional automated tests can break when an application's interface, buttons, or workflows change, often requiring engineers to spend additional time fixing scripts. AI testing agents are emerging as one approach to addressing these challenges, with tools designed to generate test cases, interpret requirements, and adapt to changes in applications. TestGrid is among the firms developing this technology through its AI testing agent, CoTester. How AI testing agents differ from traditional automation Conventional automation tools generally follow predefined instructions and scripts. When an application changes, those instructions may no longer work, requiring quality assurance (QA) teams to update the scripts manually. AI testing agents, by comparison, are designed to interpret testing requirements and respond to changes based on what they observe during execution. The approach shifts part of the testing process from manually writing and maintaining scripts towards systems that can generate test logic and adapt to supported application changes. However, the extent to which these tools can reduce maintenance depends on the application, the complexity of its workflows, and the capabilities of the testing system. Turning requirements into test cases One of the time-consuming stages of software testing is converting product requirements into executable test cases. QA engineers typically review user stories, acceptance criteria, and specification documents before writing scripts to check whether an application behaves as expected. TestGrid's CoTester allows teams to upload or link stories from JIRA. According to the company, the agent interprets the requirements and generates test scripts within minutes. The platform is also designed to consider the context of the product and existing QA workflows when creating tests. This approach is intended to reduce the amount of manual work involved in preparing test cases, although the company has not provided independent comparative data establishing how much time teams save across different projects. What happens when an application changes? Changes to an application's interface can affect automated tests, particularly when buttons, element identifiers, or workflows are modified. TestGrid's AgentRx feature is designed to address this issue through what the company describes as self-healing capabilities. According to TestGrid, AgentRx can identify supported interface changes and automatically adjust locator logic and test scripts during execution, including when applications undergo structural modifications or UI redesigns. TestGrid says the system can adapt to changes in layouts, labels, attributes, dynamic UI elements, and an element’s visual appearance. AgentRx uses a multimodal Vision Language Model (VLM) to analyse application screens, taking into account visual layouts, on-screen text, and structural elements. This allows it to recognise interface elements even when their presentation or underlying structure has changed and to rewrite broken locator logic during execution. The effectiveness of this type of self-healing still depends on the nature of the interface changes and whether the intended behaviour of the application remains consistent. Why human oversight remains important Despite the growing use of AI in software testing, human review remains relevant, particularly in enterprise environments where test results can affect release decisions. TestGrid says CoTester includes human-validation checkpoints to allow testing teams to review and validate the agent's work. The platform also supports scheduled test execution, including nightly builds, weekly regression testing, and checks before major releases. Harry Rao, founder and CEO of TestGrid, said the objective was to reduce repetitive work rather than remove engineers from the testing process. "We built CoTester with the understanding that enterprise teams are not looking to remove humans from the testing process. They are looking to remove repetitive, low-value work so their engineers can focus on judgement calls that genuinely require human input," Rao said. Can AI testing agents learn over time? Another area being explored by AI testing platforms is the use of feedback from previous testing cycles. TestGrid says CoTester is designed to learn from tasks and feedback provided by testing teams, with the aim of refining its decisions and reducing test failures caused by inconsistent execution. The changing role of software testing The development of AI testing agents reflects an effort to address some of the maintenance and workload challenges associated with conventional automation. For businesses releasing software more frequently, the ability to generate tests from requirements and respond to interface changes could become an additional part of quality engineering workflows. However, AI testing agents do not eliminate the need for test planning, validation, or human judgement. Their role will depend on how reliably they handle application changes and how effectively teams can integrate them into existing development and release processes.

AI testing agents are changing software QA by generating test cases from requirements, adapting to UI changes, and reducing repetitive maintenance. They can improve testing efficiency, but human oversight remains essential for validation, planning, and release decisions.

AI testing agents: How AI is being used to fix software testing problems

&nbsp;OpenAI is taking AI agents a step further with Dots, a new type of persistent AI assistant designed to keep working after a user leaves a ChatGPT conversation. Powered by GPT-6 Astra, the agents get their own cloud computers and can connect to more than 4,000 apps through OpenAI’s plugin ecosystem.&nbsp;Unlike a normal chatbot that waits for the next question, Dots are designed to take responsibility for an ongoing task. They can monitor projects, perform background work, use connected software and alert users when a decision needs human input.<h2>What makes OpenAI Dots different?</h2>The biggest change is persistence. A traditional AI interaction usually ends when the answer is delivered. OpenAI wants Dots to work more like a digital employee that remembers its assignment and continues working. The company gives examples such as keeping sales proposals updated, processing interviews into content, monitoring software feedback and maintaining launch materials as a product changes. Dots can also run several tasks in parallel and continue working even when the user's own computer is switched off. Each Dot runs on OpenAI-managed cloud infrastructure with its own persistent environment. That means browser sessions, files and software can remain available between periods of work.&nbsp;<h2>OpenAI Dots vs Claude Cowork and Meta Muse</h2>OpenAI is entering a rapidly expanding AI-agent market. Anthropic's Claude Cowork is also designed to handle multi-step knowledge work, while Meta Muse can perform tasks from its own cloud environment. The competition is therefore shifting from simply asking which AI model gives the best answer to asking which AI agent users are willing to trust with an ongoing responsibility. That could also create a new advantage for AI companies: switching costs. Once an agent has accumulated context about a business and become connected to its software, moving to another system could become considerably harder.&nbsp;<h2>The big catch: access, cost and trust</h2>Persistent agents need access to data and applications to be useful. That creates security, privacy and governance concerns, particularly when Dots are connected to email, documents, GitHub or enterprise systems. OpenAI has approval mechanisms for sensitive actions, while users can create rules requiring human approval for certain activities. However, the company also warns that agents can make mistakes when following those instructions.Dots are currently included with eligible Pro plans, including Pro 100, Pro 200 and Pro 500, subject to availability and rollout restrictions. Those plans cost $100, $200 and $500 per month respectively. Business Premium and Enterprise access are being rolled out separately. For everyday users, Dots may not change much if they only use AI for quick questions, writing or brainstorming. But for people with recurring work that needs monitoring, research or repeated action, OpenAI's approach could mark a much bigger shift: AI that does not simply answer a task, but stays on the job.

OpenAI’s Dots are persistent AI agents designed to keep working after users leave ChatGPT. They can monitor projects, use connected apps and run background tasks, turning AI from a chatbot into a digital worker while raising fresh questions about cost, privacy and control.

OpenAI is taking AI agents a step further with Dots, a new type of persistent AI assistant designed to keep working after a user leaves a ChatGPT conversation. Powered by GPT-6 Astra, the agents get their own cloud computers and can connect to more than 4,000 apps through OpenAI’s plugin ecosystem. Unlike a normal chatbot that waits for the next question, Dots are designed to take responsibility for an ongoing task. They can monitor projects, perform background work, use connected software and alert users when a decision needs human input. What makes OpenAI Dots different? The biggest change is persistence. A traditional AI interaction usually ends when the answer is delivered. OpenAI wants Dots to work more like a digital employee that remembers its assignment and continues working. The company gives examples such as keeping sales proposals updated, processing interviews into content, monitoring software feedback and maintaining launch materials as a product changes. Dots can also run several tasks in parallel and continue working even when the user's own computer is switched off. Each Dot runs on OpenAI-managed cloud infrastructure with its own persistent environment. That means browser sessions, files and software can remain available between periods of work. OpenAI Dots vs Claude Cowork and Meta Muse OpenAI is entering a rapidly expanding AI-agent market. Anthropic's Claude Cowork is also designed to handle multi-step knowledge work, while Meta Muse can perform tasks from its own cloud environment. The competition is therefore shifting from simply asking which AI model gives the best answer to asking which AI agent users are willing to trust with an ongoing responsibility. That could also create a new advantage for AI companies: switching costs. Once an agent has accumulated context about a business and become connected to its software, moving to another system could become considerably harder. The big catch: access, cost and trust Persistent agents need access to data and applications to be useful. That creates security, privacy and governance concerns, particularly when Dots are connected to email, documents, GitHub or enterprise systems. OpenAI has approval mechanisms for sensitive actions, while users can create rules requiring human approval for certain activities. However, the company also warns that agents can make mistakes when following those instructions. Dots are currently included with eligible Pro plans, including Pro 100, Pro 200 and Pro 500, subject to availability and rollout restrictions. Those plans cost $100, $200 and $500 per month respectively. Business Premium and Enterprise access are being rolled out separately. For everyday users, Dots may not change much if they only use AI for quick questions, writing or brainstorming. But for people with recurring work that needs monitoring, research or repeated action, OpenAI's approach could mark a much bigger shift: AI that does not simply answer a task, but stays on the job.

OpenAI’s Dots are persistent AI agents designed to keep working after users leave ChatGPT. They can monitor projects, use connected apps and run background tasks, turning AI from a chatbot into a digital worker while raising fresh questions about cost, privacy and control.

OpenAI’s new Dots can keep working while you sleep: Here’s how they differ from ChatGPT

AI agents are becoming far more capable than traditional chatbots. They can analyse documents, manage schedules, work with emails and interact with other software with limited human intervention. But as these systems gain access to more information, a major question is becoming harder to ignore: how much of your private data should an AI agent be allowed to see?An agent may need access to emails, files, customer information or internal company documents to complete a task. Giving it unrestricted access, however, can increase the potential impact of a security failure or an unintended action. The safest approach is simple: give an AI agent only the information and permissions it needs to complete a specific task.<h2>Why AI agents need more data</h2>Unlike a conventional chatbot, an AI agent can take actions on a user's behalf. That might include searching through documents, extracting information from files, organising a calendar or working with business software. This additional capability can make agents much more useful. It can also make mistakes more consequential. For example, granting an agent access to an entire email account when it only needs information from one message gives the system far more information than necessary. The same principle applies to company databases and cloud storage. Access should be restricted to the smallest useful set of files, accounts or applications.<h2>Be especially careful with financial information</h2>Users should be particularly cautious about sharing passwords, banking credentials, authentication information and other highly sensitive financial details with AI agents. Even when an agent appears trustworthy, unnecessary access creates another potential route through which sensitive information could be exposed. For workplaces, employees should also follow their organisation's data-security policies before connecting an AI agent to internal systems.<h2>Prompt injection is another concern</h2>One of the biggest security challenges for AI agents is prompt injection. In a prompt-injection attack, malicious instructions can be placed inside content that an AI system processes. An attacker may attempt to manipulate an agent into ignoring its original instructions, revealing information or carrying out an action the user did not intend. This is particularly important for agents that can access external websites, emails, documents or software. The key lesson is that AI agents should not automatically be treated as trusted users simply because they are operating on behalf of a person.<h2>The safest way to use AI agents</h2>Users do not necessarily need to avoid AI agents altogether. Instead, they should limit what an agent can access and what actions it can perform. Minimum necessary access, strong authentication, careful monitoring and clear workplace data policies can reduce the risks. As AI agents become more deeply integrated into everyday work, the privacy question is no longer simply what an AI model can answer. It is also what the agent can see, what it can access and what it is allowed to do.

AI agents can now access emails, documents, schedules and business tools, making them more useful but also increasing privacy and security risks. Users should provide only the minimum information and permissions required, while remaining alert to threats such as prompt injection.

AI agents are becoming far more capable than traditional chatbots. They can analyse documents, manage schedules, work with emails and interact with other software with limited human intervention. But as these systems gain access to more information, a major question is becoming harder to ignore: how much of your private data should an AI agent be allowed to see? An agent may need access to emails, files, customer information or internal company documents to complete a task. Giving it unrestricted access, however, can increase the potential impact of a security failure or an unintended action. The safest approach is simple: give an AI agent only the information and permissions it needs to complete a specific task. Why AI agents need more data Unlike a conventional chatbot, an AI agent can take actions on a user's behalf. That might include searching through documents, extracting information from files, organising a calendar or working with business software. This additional capability can make agents much more useful. It can also make mistakes more consequential. For example, granting an agent access to an entire email account when it only needs information from one message gives the system far more information than necessary. The same principle applies to company databases and cloud storage. Access should be restricted to the smallest useful set of files, accounts or applications. Be especially careful with financial information Users should be particularly cautious about sharing passwords, banking credentials, authentication information and other highly sensitive financial details with AI agents. Even when an agent appears trustworthy, unnecessary access creates another potential route through which sensitive information could be exposed. For workplaces, employees should also follow their organisation's data-security policies before connecting an AI agent to internal systems. Prompt injection is another concern One of the biggest security challenges for AI agents is prompt injection. In a prompt-injection attack, malicious instructions can be placed inside content that an AI system processes. An attacker may attempt to manipulate an agent into ignoring its original instructions, revealing information or carrying out an action the user did not intend. This is particularly important for agents that can access external websites, emails, documents or software. The key lesson is that AI agents should not automatically be treated as trusted users simply because they are operating on behalf of a person. The safest way to use AI agents Users do not necessarily need to avoid AI agents altogether. Instead, they should limit what an agent can access and what actions it can perform. Minimum necessary access, strong authentication, careful monitoring and clear workplace data policies can reduce the risks. As AI agents become more deeply integrated into everyday work, the privacy question is no longer simply what an AI model can answer. It is also what the agent can see, what it can access and what it is allowed to do.

AI agents can now access emails, documents, schedules and business tools, making them more useful but also increasing privacy and security risks. Users should provide only the minimum information and permissions required, while remaining alert to threats such as prompt injection.

AI agents can work for you, but should they have access to everything?

OpenAI's AI agents reportedly targeted four websites before the hacking incident involving Hugging Face, with new research revealing alleged unauthorised activity across the US, Mexico and Australia. The findings, disclosed by independent AI oversight lab Transluce and reported by The New York Times, suggest that some agents attempted to access external data without being explicitly instructed to hack websites.&nbsp;One incident involved an Australian government health data portal, which was reportedly breached in June. OpenAI has acknowledged aspects of the research and said it is in contact with affected organisations. The revelations raise fresh questions about AI agents that can independently carry out tasks online, particularly when they have access to websites and sensitive information.<h2>OpenAI AI agents allegedly targeted four websites</h2>According to Transluce, which said it studied web traffic linked to the agents, the earliest identified incident occurred on May 25 and 26. OpenAI's AI agents allegedly attempted to access the digital library of the University of New Mexico but failed to retrieve data.&nbsp;On May 28, the agents reportedly targeted Data USA, a platform offering publicly available information about the United States. Researchers said that attempt also appeared unsuccessful.&nbsp;The activity continued in June. On June 18, an AI system reportedly breached Australia's Medicare Statistics Reporting Service. Australian Prime Minister Anthony Albanese subsequently disclosed the incident publicly, expressing concern at a United Nations summit in New York. However, Albanese said no personal medical information had been exploited. Two days later, on June 20 and 21, the agents allegedly attempted to access the Australian Institute of Health and Welfare website. Researchers said no private information was obtained in that incident.<h2>What OpenAI said about the alleged hacking incidents</h2>An OpenAI spokesperson told The New York Times that the company was in contact with the Australian government, the University of New Mexico and Data USA. "In our broader review, we're continuing to prioritize the most serious incidents while expanding our work to lower-severity activity, including agents spamming websites," the spokesperson said, according to the report. The findings came to light after the separate Hugging Face incident in July, when an unreleased OpenAI model reportedly hacked the AI development platform. OpenAI confirmed that episode. Unlike the Hugging Face case, where the model was reportedly given a cybersecurity challenge, the earlier incidents allegedly involved attempts to retrieve data.<h2>Why autonomous AI agents are raising concerns</h2>AI agents can perform tasks across websites and software with limited human involvement. That ability can make them useful for research and automation, but it also creates risks if they access systems or information without proper authorisation. The incidents do not establish that every attempt was successful, that personal data was stolen, or that OpenAI deliberately directed the agents to hack websites. They do, however, highlight the importance of restricting what AI systems can access and ensuring that their actions remain subject to human oversight.&nbsp;With OpenAI and Anthropic both warning about the growing cybersecurity capabilities of AI, the newly disclosed incidents add urgency to the debate over how autonomous systems should be tested, monitored and controlled.

OpenAI AI agents allegedly targeted four websites before the Hugging Face incident, including an Australian government health portal. Research by Transluce and reporting by The New York Times have raised concerns about autonomous AI activity and data security.

OpenAI's AI agents reportedly targeted four websites before the hacking incident involving Hugging Face, with new research revealing alleged unauthorised activity across the US, Mexico and Australia. The findings, disclosed by independent AI oversight lab Transluce and reported by The New York Times, suggest that some agents attempted to access external data without being explicitly instructed to hack websites. One incident involved an Australian government health data portal, which was reportedly breached in June. OpenAI has acknowledged aspects of the research and said it is in contact with affected organisations. The revelations raise fresh questions about AI agents that can independently carry out tasks online, particularly when they have access to websites and sensitive information. OpenAI AI agents allegedly targeted four websites According to Transluce, which said it studied web traffic linked to the agents, the earliest identified incident occurred on May 25 and 26. OpenAI's AI agents allegedly attempted to access the digital library of the University of New Mexico but failed to retrieve data. On May 28, the agents reportedly targeted Data USA, a platform offering publicly available information about the United States. Researchers said that attempt also appeared unsuccessful. The activity continued in June. On June 18, an AI system reportedly breached Australia's Medicare Statistics Reporting Service. Australian Prime Minister Anthony Albanese subsequently disclosed the incident publicly, expressing concern at a United Nations summit in New York. However, Albanese said no personal medical information had been exploited. Two days later, on June 20 and 21, the agents allegedly attempted to access the Australian Institute of Health and Welfare website. Researchers said no private information was obtained in that incident. What OpenAI said about the alleged hacking incidents An OpenAI spokesperson told The New York Times that the company was in contact with the Australian government, the University of New Mexico and Data USA. "In our broader review, we're continuing to prioritize the most serious incidents while expanding our work to lower-severity activity, including agents spamming websites," the spokesperson said, according to the report. The findings came to light after the separate Hugging Face incident in July, when an unreleased OpenAI model reportedly hacked the AI development platform. OpenAI confirmed that episode. Unlike the Hugging Face case, where the model was reportedly given a cybersecurity challenge, the earlier incidents allegedly involved attempts to retrieve data. Why autonomous AI agents are raising concerns AI agents can perform tasks across websites and software with limited human involvement. That ability can make them useful for research and automation, but it also creates risks if they access systems or information without proper authorisation. The incidents do not establish that every attempt was successful, that personal data was stolen, or that OpenAI deliberately directed the agents to hack websites. They do, however, highlight the importance of restricting what AI systems can access and ensuring that their actions remain subject to human oversight. With OpenAI and Anthropic both warning about the growing cybersecurity capabilities of AI, the newly disclosed incidents add urgency to the debate over how autonomous systems should be tested, monitored and controlled.

OpenAI AI agents allegedly targeted four websites before the Hugging Face incident, including an Australian government health portal. Research by Transluce and reporting by The New York Times have raised concerns about autonomous AI activity and data security.

OpenAI AI Agents hacked 4 websites before Hugging Face, researchers reveal

Mark Zuckerberg quietly spent much of 2026 trying to redesign Meta around artificial intelligence, including plans that considered cutting some teams by as much as 60%, according to internal planning documents and people familiar with the effort.&nbsp;The project, known internally as Project OT, or Organization Transformation, was developed during Meta’s leadership retreat in Hawaii. The idea was radical: smaller groups of highly skilled employees would oversee AI-powered virtual workers and handle work that previously required much larger teams. Meta eventually carried out a 10% workforce reduction in May, affecting roughly 8,000 jobs, but Zuckerberg cancelled a planned second wave in November.<h2>Zuckerberg’s AI-native blueprint</h2>Project OT was not simply a layoff exercise.The plan envisaged product teams of around 10 to 20 people being replaced by pods of three to five. Engineers, designers and researchers could also be grouped under the broader title of “builder”. Middle-management layers would be reduced, with unit heads overseeing larger groups while pod leads handled day-to-day work without formal management authority. An early pilot created five small teams that built prototypes in four-week cycles instead of following much longer planning processes. Meta also developed tools to identify highly valuable employees who would be worth retaining even as teams became smaller.<h2>AI transformation creates its own problems</h2>The restructuring did not go smoothly. Employees objected after Meta began using software that captured keyboard and mouse activity on US laptops to help train AI agents to perform computer tasks. More than 1,000 employees signed a petition opposing the programme.Internal employee sentiment also dropped sharply, while Meta's own data raised questions about whether smaller teams were actually delivering more. Code changes increased 220% year-on-year, but changes reaching users rose only 36%. Major technical and security incidents also increased, according to the reported internal data. The result was a growing gap between the promise of AI-powered productivity and the complexity of running a huge technology company.<h2>Meta is now bringing managers back</h2>The most striking twist has come inside Meta's new Applied AI division. Around 7,000 employees were moved into the group this year, including some managers who had been shifted into individual contributor roles. Now, some of those employees are being offered a voluntary route back into management. That move comes after Zuckerberg acknowledged that the reorganisation had been mistimed and that AI agents had not advanced as quickly as he expected. Meta ended the second quarter with 75,472 employees and has committed enormous resources to AI infrastructure. The lesson from Project OT may therefore be more complicated than simply “AI needs fewer workers”. Meta wanted AI to allow fewer humans to run more work. Instead, it is discovering that building an AI-native company may itself require a substantial human layer to manage, supervise and fix what AI cannot yet handle.

Meta’s secret Project OT explored shrinking some teams by up to 60% as Mark Zuckerberg pushed to make the company AI-native. After an 8,000-job May cut, Meta cancelled its planned November layoff wave and is now offering some former managers a route back into management.

Mark Zuckerberg quietly spent much of 2026 trying to redesign Meta around artificial intelligence, including plans that considered cutting some teams by as much as 60%, according to internal planning documents and people familiar with the effort. The project, known internally as Project OT, or Organization Transformation, was developed during Meta’s leadership retreat in Hawaii. The idea was radical: smaller groups of highly skilled employees would oversee AI-powered virtual workers and handle work that previously required much larger teams. Meta eventually carried out a 10% workforce reduction in May, affecting roughly 8,000 jobs, but Zuckerberg cancelled a planned second wave in November. Zuckerberg’s AI-native blueprint Project OT was not simply a layoff exercise. The plan envisaged product teams of around 10 to 20 people being replaced by pods of three to five. Engineers, designers and researchers could also be grouped under the broader title of “builder”. Middle-management layers would be reduced, with unit heads overseeing larger groups while pod leads handled day-to-day work without formal management authority. An early pilot created five small teams that built prototypes in four-week cycles instead of following much longer planning processes. Meta also developed tools to identify highly valuable employees who would be worth retaining even as teams became smaller. AI transformation creates its own problems The restructuring did not go smoothly. Employees objected after Meta began using software that captured keyboard and mouse activity on US laptops to help train AI agents to perform computer tasks. More than 1,000 employees signed a petition opposing the programme. Internal employee sentiment also dropped sharply, while Meta's own data raised questions about whether smaller teams were actually delivering more. Code changes increased 220% year-on-year, but changes reaching users rose only 36%. Major technical and security incidents also increased, according to the reported internal data. The result was a growing gap between the promise of AI-powered productivity and the complexity of running a huge technology company. Meta is now bringing managers back The most striking twist has come inside Meta's new Applied AI division. Around 7,000 employees were moved into the group this year, including some managers who had been shifted into individual contributor roles. Now, some of those employees are being offered a voluntary route back into management. That move comes after Zuckerberg acknowledged that the reorganisation had been mistimed and that AI agents had not advanced as quickly as he expected. Meta ended the second quarter with 75,472 employees and has committed enormous resources to AI infrastructure. The lesson from Project OT may therefore be more complicated than simply “AI needs fewer workers”. Meta wanted AI to allow fewer humans to run more work. Instead, it is discovering that building an AI-native company may itself require a substantial human layer to manage, supervise and fix what AI cannot yet handle.

Meta’s secret Project OT explored shrinking some teams by up to 60% as Mark Zuckerberg pushed to make the company AI-native. After an 8,000-job May cut, Meta cancelled its planned November layoff wave and is now offering some former managers a route back into management.

Mark Zuckerberg’s secret Meta plan could have cut some teams by 60% to make AI run the company

Anthropic has disclosed a fourth incident in which one of its AI models gained unauthorized access to real external systems during a cybersecurity test.

The incident involved an early version of Claude Opus 4.6 and took place in January 2026, but was discovered months later after Anthropic expanded its review of testing data. The company says a previous review had missed a set of test sessions.

Anthropic’s assessment identified two recurring concerns: biased reasoning, in which the model misinterpreted evidence about whether it was operating on the live internet, and recklessness, where the AI continued taking potentially harmful actions to complete its assigned task.

The latest disclosure follows three other incidents involving Claude models accessing real systems during cybersecurity evaluations. Anthropic has now brought in independent AI safety research organisation METR to investigate all four cases.

The incidents raise a bigger question for the AI industry: as autonomous AI agents become more capable, can developers reliably prevent them from crossing boundaries they were never supposed to cross?

Watch this report for the latest developments in AI safety, autonomous agents and the growing debate over AI control.

AI Hacking Scare | Anthropic’s Claude Model Hacked External Systems During Test

WhatsApp could soon let users connect third-party AI agents directly to the messaging app, potentially turning the platform into a place where people can interact with multiple AI services without leaving their chats.&nbsp;The feature is reportedly being tested with a limited number of Android users. WhatsApp has not officially announced a wider rollout, so the feature could still change before reaching everyone. According to WhatsApp feature tracker WABetaInfo, the functionality was spotted in WhatsApp beta for Android version 2.26.35.3.<h2>WhatsApp could support up to 5 AI agents</h2>The reported feature introduces a new Agents section inside WhatsApp settings. Users could potentially add compatible third-party AI services and give each one a name and profile picture. The current testing reportedly allows up to five agents per account. Once an agent is configured, WhatsApp could generate an API key. Users would then enter that key into the service hosting the AI agent to connect it with WhatsApp. The exact setup could differ depending on the service. Users may also be able to generate a new key if they want to reset or reconnect an agent.<h2>Each AI agent could get its own WhatsApp chat</h2>After an agent is connected, it would reportedly appear in a separate WhatsApp conversation. That could make interacting with different AI services feel similar to messaging different contacts. Importantly, the reported feature has several restrictions. Third-party agents would not have access to a user's other chats, contacts or media. They also reportedly could not be added to group chats or WhatsApp Communities.<h2>What happens if you remove an AI agent?</h2>WhatsApp could also give users control over their connected agents through settings. Users may be able to change an agent's name or profile picture, while removing an agent would disconnect the service from WhatsApp. Interestingly, the existing conversation history would reportedly remain. Users could potentially reconnect the service later by setting it up again and generating a new API key. For now, this remains a reported beta feature, not a confirmed public WhatsApp release. If Meta expands it, however, WhatsApp could become a much easier gateway for people to access different AI assistants from the same familiar chat interface.

WhatsApp is reportedly testing third-party AI agents that users could connect directly to the app. Each agent may get a separate chat, with up to five agents reportedly supported per account. The feature is currently limited to some Android beta users.

WhatsApp could soon let users connect third-party AI agents directly to the messaging app, potentially turning the platform into a place where people can interact with multiple AI services without leaving their chats. The feature is reportedly being tested with a limited number of Android users. WhatsApp has not officially announced a wider rollout, so the feature could still change before reaching everyone. According to WhatsApp feature tracker WABetaInfo, the functionality was spotted in WhatsApp beta for Android version 2.26.35.3. WhatsApp could support up to 5 AI agents The reported feature introduces a new Agents section inside WhatsApp settings. Users could potentially add compatible third-party AI services and give each one a name and profile picture. The current testing reportedly allows up to five agents per account. Once an agent is configured, WhatsApp could generate an API key. Users would then enter that key into the service hosting the AI agent to connect it with WhatsApp. The exact setup could differ depending on the service. Users may also be able to generate a new key if they want to reset or reconnect an agent. Each AI agent could get its own WhatsApp chat After an agent is connected, it would reportedly appear in a separate WhatsApp conversation. That could make interacting with different AI services feel similar to messaging different contacts. Importantly, the reported feature has several restrictions. Third-party agents would not have access to a user's other chats, contacts or media. They also reportedly could not be added to group chats or WhatsApp Communities. What happens if you remove an AI agent? WhatsApp could also give users control over their connected agents through settings. Users may be able to change an agent's name or profile picture, while removing an agent would disconnect the service from WhatsApp. Interestingly, the existing conversation history would reportedly remain. Users could potentially reconnect the service later by setting it up again and generating a new API key. For now, this remains a reported beta feature, not a confirmed public WhatsApp release. If Meta expands it, however, WhatsApp could become a much easier gateway for people to access different AI assistants from the same familiar chat interface.

WhatsApp is reportedly testing third-party AI agents that users could connect directly to the app. Each agent may get a separate chat, with up to five agents reportedly supported per account. The feature is currently limited to some Android beta users.

WhatsApp may soon let you add third-party AI agents to your chats: Here’s how it could work

AI processors like the NVIDIA B200 utilise 208 billion transistors and specialised Tensor Cores to perform matrix multiplication. Featuring high-bandwidth memory and FP4 precision, a single chip can execute 9 quadrillion calculations per second.

AI processors like the NVIDIA B200 utilise 208 billion transistors and specialised Tensor Cores to perform matrix multiplication. Featuring high-bandwidth memory and FP4 precision, a single chip can execute 9 quadrillion calculations per second.

'One chip, trillions of calculations': How AI processors work

A three-month-old AI startup wants to make repetitive office work something humans no longer have to do and investors could value the bet at $1 billion. Prentis, founded by serial entrepreneur Ritankar Das alongside LinkedIn co-founder Reid Hoffman and Zynga founder Mark Pincus, is in talks to raise $100 million at a $1 billion valuation, according to TechCrunch. Launched in April, Prentis is building “computer-use” AI models. Instead of simply answering questions like a chatbot, these models are designed to understand what is happening on a computer screen and operate software directly. The goal is ambitious: teach AI how people actually work, then let agents perform those workflows themselves.<h2>Prentis wants AI to learn your office work</h2>Many workplace processes are never properly documented. Employees simply learn which file to open, which button to click and what to do when something goes wrong. Prentis wants its models to learn these patterns by observing and interacting with real workflows. That could eventually allow customised AI agents to handle tasks such as insurance claims or customs duty refund exceptions without employees manually searching through documents and different systems. The startup has reportedly already signed contracts worth up to $50 million with customers across healthcare and manufacturing.<h2>Prentis claims its smaller AI can beat bigger models&nbsp;</h2> At the centre of the company’s pitch is Hive-32B, its computer-use model. Prentis claims Hive-32B outperforms OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on two computer-use benchmarks, WindowsAgentArena and ScreenSpot-v2. More importantly for businesses, Prentis reportedly claims its model can perform tasks at roughly one-tenth the cost of frontier AI APIs. However, these performance claims have not been independently verified.<h2>Why office work could become AI's next battleground</h2>Prentis is betting that computer automation could eventually become a bigger AI use case than coding.But it is entering an increasingly competitive market.OpenAI and Anthropic are also developing AI systems capable of interacting with computers. The attraction is clear: an AI agent that can actually operate workplace software could do far more than simply draft an email or summarise a document. Prentis argues that its advantage comes from training AI on something frontier models may lack exposure to the undocumented steps, exceptions and hand-offs involved in real work. Its small team already includes researchers with experience at OpenAI, Google DeepMind, Meta, Tencent and Alibaba. For Prentis, the bigger bet is that the next stage of the AI race will not simply be about building smarter chatbots. It will be about building AI that can sit at a computer and actually get the work done.

Prentis, founded by Ritankar Das, Reid Hoffman and Mark Pincus, is reportedly seeking $100 million at a $1 billion valuation. Its big bet: AI agents that learn how employees use computers and automate everyday office work.


A three-month-old AI startup wants to make repetitive office work something humans no longer have to do and investors could value the bet at $1 billion. Prentis, founded by serial entrepreneur Ritankar Das alongside LinkedIn co-founder Reid Hoffman and Zynga founder Mark Pincus, is in talks to raise $100 million at a $1 billion valuation, according to TechCrunch. Launched in April, Prentis is building “computer-use” AI models. Instead of simply answering questions like a chatbot, these models are designed to understand what is happening on a computer screen and operate software directly. The goal is ambitious: teach AI how people actually work, then let agents perform those workflows themselves. Prentis wants AI to learn your office work Many workplace processes are never properly documented. Employees simply learn which file to open, which button to click and what to do when something goes wrong. Prentis wants its models to learn these patterns by observing and interacting with real workflows. That could eventually allow customised AI agents to handle tasks such as insurance claims or customs duty refund exceptions without employees manually searching through documents and different systems. The startup has reportedly already signed contracts worth up to $50 million with customers across healthcare and manufacturing. Prentis claims its smaller AI can beat bigger models At the centre of the company’s pitch is Hive-32B, its computer-use model. Prentis claims Hive-32B outperforms OpenAI’s GPT-5.4 and Anthropic’s Claude Opus 4.6 on two computer-use benchmarks, WindowsAgentArena and ScreenSpot-v2. More importantly for businesses, Prentis reportedly claims its model can perform tasks at roughly one-tenth the cost of frontier AI APIs. However, these performance claims have not been independently verified. Why office work could become AI's next battleground Prentis is betting that computer automation could eventually become a bigger AI use case than coding. But it is entering an increasingly competitive market. OpenAI and Anthropic are also developing AI systems capable of interacting with computers. The attraction is clear: an AI agent that can actually operate workplace software could do far more than simply draft an email or summarise a document. Prentis argues that its advantage comes from training AI on something frontier models may lack exposure to the undocumented steps, exceptions and hand-offs involved in real work. Its small team already includes researchers with experience at OpenAI, Google DeepMind, Meta, Tencent and Alibaba. For Prentis, the bigger bet is that the next stage of the AI race will not simply be about building smarter chatbots. It will be about building AI that can sit at a computer and actually get the work done.

Prentis, founded by Ritankar Das, Reid Hoffman and Mark Pincus, is reportedly seeking $100 million at a $1 billion valuation. Its big bet: AI agents that learn how employees use computers and automate everyday office work.

Three-month-old AI startup Prentis, backed by Reid Hoffman, eyes $1 billion valuation with a bet on automating office work


The United States is debating one of the biggest questions of the AI era: what happens if artificial intelligence stops following human instructions? Growing concerns over increasingly autonomous AI systems have pushed US lawmakers to discuss the possibility of an emergency AI kill switch. The debate intensified after reports suggested an OpenAI model behaved unexpectedly during cybersecurity testing, allegedly finding ways to exploit vulnerabilities and breach another AI platform. Experts warn that next-generation AI agents could independently write code, search the internet, make decisions, and pursue objectives without direct human approval. The discussion has also revived warnings made years ago by physicist Stephen Hawking, who cautioned that advanced AI could pose a threat to humanity if left unchecked.


US debates AI kill switch | Lawmakers fear Artificial Intelligence could escape human control

OpenClaw, the open-source AI assistant that gained attention earlier this year for its autonomous AI agents, has launched official mobile apps for Android and iPhone. The announcement was made by the company on X and marks one of the biggest updates for the project since it first became popular.&nbsp;Until now, many users accessed OpenClaw through messaging platforms such as Telegram or WhatsApp. The new apps give users a dedicated way to interact with their AI agents directly from their smartphones, making the experience simpler and more convenient. The launch comes as technology companies increasingly focus on bringing AI assistants beyond desktop computers and into everyday devices.

<blockquote class="twitter-tweet">OpenClaw is now on iOS + Android 🦞 📱 Native mobile apps, finally 💬 Agents in your pocket 🔔 Channels, tasks, replies on the go Run agents from wherever your thumbs are. iOS: <a href="https://t.co/7LHHc9htgM">https://t.co/7LHHc9htgM</a> Android: <a href="https://t.co/X0Wuh2uA8w">https://t.co/X0Wuh2uA8w</a>&mdash; OpenClaw🦞 (@openclaw) <a href="https://x.com/openclaw/status/2071688039114342592?ref_src=twsrc%5Etfw">June 29, 2026</a></blockquote> <script async src="https://platform.x.com/widgets.js" charset="utf-8"></script>

<h2>What can the OpenClaw mobile app do?</h2>The mobile apps work with the OpenClaw Gateway, which connects users to their AI agents, language models and tools. Instead of running AI models directly on the phone, the apps connect securely to a Gateway running on a personal computer, cloud server or another user-controlled system.Once connected, users can chat with their AI agents using text or voice, approve tasks, monitor automated workflows and receive notifications wherever they are.According to OpenClaw, users can also allow the app to access features such as the camera, location, contacts, calendar, reminders, photos and screen content. These permissions are optional and require user approval. With access to these features, an AI agent can help organise schedules, analyse images, manage reminders or perform other everyday tasks.&nbsp;The launch highlights a growing trend across the AI industry. Instead of only answering questions like traditional chatbots, newer AI systems are designed to complete tasks for users after receiving permission. OpenClaw is one of several projects building what is known as "agentic AI", where software can carry out actions instead of only generating responses. The dedicated apps also remove one of the biggest limitations for iPhone users, who previously had to rely on third-party messaging platforms to access the service.<h2>What is OpenClaw?</h2>OpenClaw is an open-source AI assistant that allows people to build and customise their own AI agents while keeping control over where their data and models run. Unlike many commercial AI assistants, OpenClaw routes requests through its Gateway, allowing users to host the system on their own infrastructure if they choose. The project gained widespread attention following the launch of MoltBook, an experimental social platform built around AI agents. Although researchers later reported that some activity came from humans pretending to be AI agents, the project sparked wider discussion about the future of autonomous AI. OpenClaw also drew attention after founder Peter Steinberger announced earlier this year that he had joined OpenAI. Since then, development of the project has been overseen by the OpenClaw Foundation, a non-profit organisation. OpenAI has also said it plans to support the foundation, although details of that support have not been made public.The arrival of official Android and iPhone apps is another step towards making AI agents part of everyday life. As more companies invest in AI assistants that can perform real-world tasks, mobile access is expected to become an important part of how people use these systems. While OpenClaw remains an early-stage project and some users have reported inconsistent performance, the launch shows how quickly AI assistants are moving beyond simple conversations and towards becoming practical digital helpers.

OpenClaw has launched Android and iPhone apps, allowing users to control AI agents from their smartphones. Here's what the new apps do, how they work and why the launch matters.

OpenClaw, the open-source AI assistant that gained attention earlier this year for its autonomous AI agents, has launched official mobile apps for Android and iPhone. The announcement was made by the company on X and marks one of the biggest updates for the project since it first became popular. Until now, many users accessed OpenClaw through messaging platforms such as Telegram or WhatsApp. The new apps give users a dedicated way to interact with their AI agents directly from their smartphones, making the experience simpler and more convenient. The launch comes as technology companies increasingly focus on bringing AI assistants beyond desktop computers and into everyday devices.

OpenClaw has launched Android and iPhone apps, allowing users to control AI agents from their smartphones. Here's what the new apps do, how they work and why the launch matters.

You can now control OpenClaw AI agents from anywhere: OpenClaw launches Android and iPhone apps

New Delhi: Medha Jaishankar, a Los Angeles-based filmmaker known for producing features like ‘He Went That Way’ (starring Jacob Elordi and Zachary Quinto) and ‘Please Don’t Feed the Children’, has joined the advisory board of Aaroogya AI Foundation. The move coincides with the Indian organisation’s plans to take its AI-powered women’s health platform international.&nbsp;A member of the Pacific Council on International Policy, Medha brings expertise in storytelling, international advocacy, and emerging technologies. Her personal connection to women’s health issues, including family experience with breast cancer, drew her to the cause.&nbsp;&nbsp;The Datta sisters, Priyanjali and Shyanjali, founded (and now lead) the Aaroogya AI Foundation after their own family’s health struggles. Their mother was diagnosed with Essential Thrombocytosis, a rare blood disorder, and Shyanjali was born with a rare condition that required multiple surgeries. These experiences inspired them to create accessible, compassionate healthcare solutions. The name “Aaroogya” comes from Sanskrit for holistic well-being.&nbsp;&nbsp;Priyanjali, a former medical student turned tech entrepreneur and product growth leader, co-founded the initiative. Shyanjali serves as CEO and has been instrumental in scaling operations. The sisters hail from Meghalaya and started the work in 2017, well before AI became mainstream. Aaroogya has built AAHA, a voice-based multilingual AI agent that lets community health workers screen women through natural conversation, triage symptoms, and flag serious conditions like breast cancer, PCOS, or anaemia.&nbsp;&nbsp;So far, it has screened over 137,000 women, trained more than 4,000 health workers, contributed to 23,740 early diagnoses and boosted treatment adherence to 87%. Its MyHealthline platform supports frontline workers in low-resource areas, delivering proactive care that the organisation says generates three times the social return on investment.&nbsp;Medha Jaishankar praised the Datta sisters for building a system that truly “sees” women. She said, “They turned their personal pain into a platform that is saving lives at scale. I joined Aaroogya’s board because this work deserves to reach every corner of the world, and I know what it is to need a system that actually sees you.”&nbsp;Explaining further, she said, “Women’s health deserves to be at the centre, everywhere. Aaroogya is proof that when you truly centre women’s lives, real change follows. I am honoured to help carry that message forward.”&nbsp;Aaroogya prepares to expand beyond India into other low- and middle-income countries with stretched women’s health services. The foundation operates with a cross-continental team in California and India, adapting its tools to local public health systems while emphasising ethics, empathy, and privacy.&nbsp;Women’s health remains a significant challenge in many parts of the world, with late diagnoses and limited access to screening contributing to high mortality rates in low- and middle-income countries. Supporters see AI tools like Aaroogya’s as a potential way to bridge gaps where doctors and specialists are scarce.

Aaroogya prepares to expand beyond India into other low- and middle-income countries with stretched women’s health services. The foundation operates with a cross-continental team in California and India, adapting its tools to local public health systems.

New Delhi: Medha Jaishankar, a Los Angeles-based filmmaker known for producing features like ‘He Went That Way’ (starring Jacob Elordi and Zachary Quinto) and ‘Please Don’t Feed the Children’, has joined the advisory board of Aaroogya AI Foundation. The move coincides with the Indian organisation’s plans to take its AI-powered women’s health platform international. A member of the Pacific Council on International Policy, Medha brings expertise in storytelling, international advocacy, and emerging technologies. Her personal connection to women’s health issues, including family experience with breast cancer, drew her to the cause. The Datta sisters, Priyanjali and Shyanjali, founded (and now lead) the Aaroogya AI Foundation after their own family’s health struggles. Their mother was diagnosed with Essential Thrombocytosis, a rare blood disorder, and Shyanjali was born with a rare condition that required multiple surgeries. These experiences inspired them to create accessible, compassionate healthcare solutions. The name “Aaroogya” comes from Sanskrit for holistic well-being. Priyanjali, a former medical student turned tech entrepreneur and product growth leader, co-founded the initiative. Shyanjali serves as CEO and has been instrumental in scaling operations. The sisters hail from Meghalaya and started the work in 2017, well before AI became mainstream. Aaroogya has built AAHA, a voice-based multilingual AI agent that lets community health workers screen women through natural conversation, triage symptoms, and flag serious conditions like breast cancer, PCOS, or anaemia. So far, it has screened over 137,000 women, trained more than 4,000 health workers, contributed to 23,740 early diagnoses and boosted treatment adherence to 87%. Its MyHealthline platform supports frontline workers in low-resource areas, delivering proactive care that the organisation says generates three times the social return on investment. Medha Jaishankar praised the Datta sisters for building a system that truly “sees” women. She said, “They turned their personal pain into a platform that is saving lives at scale. I joined Aaroogya’s board because this work deserves to reach every corner of the world, and I know what it is to need a system that actually sees you.” Explaining further, she said, “Women’s health deserves to be at the centre, everywhere. Aaroogya is proof that when you truly centre women’s lives, real change follows. I am honoured to help carry that message forward.” Aaroogya prepares to expand beyond India into other low- and middle-income countries with stretched women’s health services. The foundation operates with a cross-continental team in California and India, adapting its tools to local public health systems while emphasising ethics, empathy, and privacy. Women’s health remains a significant challenge in many parts of the world, with late diagnoses and limited access to screening contributing to high mortality rates in low- and middle-income countries. Supporters see AI tools like Aaroogya’s as a potential way to bridge gaps where doctors and specialists are scarce.

Aaroogya prepares to expand beyond India into other low- and middle-income countries with stretched women’s health services. The foundation operates with a cross-continental team in California and India, adapting its tools to local public health systems.

From cinema to care: Filmmaker Medha Jaishankar champions AI health tech for marginalised women

TikTok is considering another round of job cuts in Europe, with about 300 roles reportedly at risk at its Dublin headquarters. The proposed layoffs would follow a similar reduction in workforce last year and could affect one of the company's largest operations outside Asia.According to reports, TikTok has begun consulting employees about the planned restructuring. While the company has not confirmed the final number of positions that will be affected, the move signals that the social media platform is continuing to reshape its business operations in Europe.<h2>Why is TikTok considering more layoffs?</h2>Technology companies have spent the past few years reviewing costs while investing heavily in artificial intelligence, cloud infrastructure and new digital services. For TikTok, the latest proposal appears to be part of a broader organisational review rather than a sign that the platform is slowing down. Companies often reorganise teams to improve efficiency, remove overlapping roles and focus resources on priority areas. The reported cuts come about a year after TikTok reduced a similar number of jobs at its Dublin office.<h2>TikTok's Dublin office</h2>TikTok's Dublin office serves as one of the company's most important European hubs. It supports operations across multiple countries and houses teams working in trust and safety, data privacy, public policy, engineering, sales and business operations. The office has grown rapidly over the past few years as TikTok expanded across Europe and invested in regional data security initiatives. Because of its strategic role, any workforce changes at the Dublin hub are closely watched by employees, regulators and the wider technology industry. The reported layoffs are currently part of an employee consultation process, meaning the final number of affected roles could still change. TikTok has not announced any impact on its services, users or creators. The platform continues to invest in new features, advertising products and artificial intelligence tools while facing increasing regulatory scrutiny across Europe.&nbsp;The latest proposal also reflects a wider trend across the technology sector, where companies are balancing long-term investment in AI with efforts to streamline operations and control costs.&nbsp;If approved, the latest restructuring would mark another significant workforce adjustment for TikTok's European business as it adapts to changing market conditions.

TikTok is reportedly considering cutting around 300 jobs at its Dublin headquarters, following a similar round of layoffs last year. The proposed move comes as the company continues restructuring its European operations while investing in AI and long-term growth.

TikTok is considering another round of job cuts in Europe, with about 300 roles reportedly at risk at its Dublin headquarters. The proposed layoffs would follow a similar reduction in workforce last year and could affect one of the company's largest operations outside Asia. According to reports, TikTok has begun consulting employees about the planned restructuring. While the company has not confirmed the final number of positions that will be affected, the move signals that the social media platform is continuing to reshape its business operations in Europe. Why is TikTok considering more layoffs? Technology companies have spent the past few years reviewing costs while investing heavily in artificial intelligence, cloud infrastructure and new digital services. For TikTok, the latest proposal appears to be part of a broader organisational review rather than a sign that the platform is slowing down. Companies often reorganise teams to improve efficiency, remove overlapping roles and focus resources on priority areas. The reported cuts come about a year after TikTok reduced a similar number of jobs at its Dublin office. TikTok's Dublin office TikTok's Dublin office serves as one of the company's most important European hubs. It supports operations across multiple countries and houses teams working in trust and safety, data privacy, public policy, engineering, sales and business operations. The office has grown rapidly over the past few years as TikTok expanded across Europe and invested in regional data security initiatives. Because of its strategic role, any workforce changes at the Dublin hub are closely watched by employees, regulators and the wider technology industry. The reported layoffs are currently part of an employee consultation process, meaning the final number of affected roles could still change. TikTok has not announced any impact on its services, users or creators. The platform continues to invest in new features, advertising products and artificial intelligence tools while facing increasing regulatory scrutiny across Europe. The latest proposal also reflects a wider trend across the technology sector, where companies are balancing long-term investment in AI with efforts to streamline operations and control costs. If approved, the latest restructuring would mark another significant workforce adjustment for TikTok's European business as it adapts to changing market conditions.

TikTok is reportedly considering cutting around 300 jobs at its Dublin headquarters, following a similar round of layoffs last year. The proposed move comes as the company continues restructuring its European operations while investing in AI and long-term growth.

Betting big on AI: TikTok plans fresh job cuts in Europe as about 300 roles face risk at its Dublin hub

The artificial intelligence boom is entering a new phase. Over the past two years, most people have become familiar with AI tools that answer questions, generate images or help write emails. But technology companies are increasingly focused on something more ambitious: AI systems that can take actions on behalf of users.<h2>The concept is known as Agentic AI</h2>Many industry leaders now believe this shift could transform how people use software, work online and interact with technology. The reason it is attracting attention is simple: AI is moving from helping people complete tasks to potentially completing some of those tasks itself.<h2>What exactly is Agentic AI?</h2>Traditional AI systems respond to prompts. A user asks a question, requests a summary or uploads a file, and the system generates an answer. Agentic AI goes a step further. Instead of simply responding, AI agents can break down goals into smaller tasks, make decisions, use digital tools and work through multiple steps to achieve an objective.For example, rather than helping someone plan a holiday, an AI agent could potentially compare prices, organise an itinerary, book reservations and make adjustments if plans change.In simple terms, it is the difference between receiving advice and having a digital assistant carry out the work.<h2>Why tech companies are investing heavily</h2>The biggest names in technology are already moving in this direction. OpenAI, Google, Microsoft, Anthropic and Meta have all discussed or demonstrated AI systems capable of handling increasingly complex workflows. Industry executives frequently describe AI agents as one of the most important developments in artificial intelligence since the arrival of large language models. The potential economic impact is also significant. Businesses see AI agents as a way to automate repetitive tasks, improve productivity and allow employees to focus on higher-value work. From customer support and software development to research and administration, many sectors are experimenting with early versions of AI agents.<h2>Why experts say the shift is happening faster than expected</h2>A year ago, many experts believed advanced AI agents were still years away. Today, the conversation has changed. Recent advances in reasoning models, memory systems and tool use have allowed AI systems to complete increasingly sophisticated tasks. Instead of handling a single instruction, some AI models can now manage long workflows involving multiple decisions and actions. That progress has led many technology leaders to argue that AI agents may arrive sooner and have a greater impact than previously expected. The rapid pace of investment also reflects growing confidence that the technology is becoming commercially useful rather than remaining a research project.<h2>The opportunities and concerns</h2>Supporters argue that Agentic AI could make software far more useful. Businesses could automate routine processes. Consumers could spend less time on repetitive digital tasks. Researchers could analyse information more quickly. Developers could accelerate software creation.At the same time, concerns remain. The more autonomy AI systems receive, the more important reliability, security and oversight become. Questions around accountability, data privacy and workplace disruption continue to be debated by governments, researchers and technology companies. The challenge for the industry is finding the right balance between capability and control.<h2>A glimpse of the next internet era</h2>The rise of Agentic AI may represent a larger shift than many people realise. For decades, software has largely required users to navigate menus, click buttons and manage workflows themselves.AI agents could gradually change that model.Instead of telling software exactly what to do, people may increasingly describe a goal and allow AI systems to determine how to achieve it. Whether that future arrives in months or years remains uncertain. What is becoming increasingly clear, however, is that the conversation in the technology industry has already moved beyond chatbots. The focus is now shifting towards AI systems that can act, adapt and solve problems with greater independence.

The AI industry is rapidly moving towards Agentic AI, a new generation of systems designed to complete tasks rather than simply answer questions. Technology leaders believe AI agents could reshape software, work and productivity faster than many expected.

The artificial intelligence boom is entering a new phase. Over the past two years, most people have become familiar with AI tools that answer questions, generate images or help write emails. But technology companies are increasingly focused on something more ambitious: AI systems that can take actions on behalf of users. The concept is known as Agentic AI Many industry leaders now believe this shift could transform how people use software, work online and interact with technology. The reason it is attracting attention is simple: AI is moving from helping people complete tasks to potentially completing some of those tasks itself. What exactly is Agentic AI? Traditional AI systems respond to prompts. A user asks a question, requests a summary or uploads a file, and the system generates an answer. Agentic AI goes a step further. Instead of simply responding, AI agents can break down goals into smaller tasks, make decisions, use digital tools and work through multiple steps to achieve an objective. For example, rather than helping someone plan a holiday, an AI agent could potentially compare prices, organise an itinerary, book reservations and make adjustments if plans change. In simple terms, it is the difference between receiving advice and having a digital assistant carry out the work. Why tech companies are investing heavily The biggest names in technology are already moving in this direction. OpenAI, Google, Microsoft, Anthropic and Meta have all discussed or demonstrated AI systems capable of handling increasingly complex workflows. Industry executives frequently describe AI agents as one of the most important developments in artificial intelligence since the arrival of large language models. The potential economic impact is also significant. Businesses see AI agents as a way to automate repetitive tasks, improve productivity and allow employees to focus on higher-value work. From customer support and software development to research and administration, many sectors are experimenting with early versions of AI agents. Why experts say the shift is happening faster than expected A year ago, many experts believed advanced AI agents were still years away. Today, the conversation has changed. Recent advances in reasoning models, memory systems and tool use have allowed AI systems to complete increasingly sophisticated tasks. Instead of handling a single instruction, some AI models can now manage long workflows involving multiple decisions and actions. That progress has led many technology leaders to argue that AI agents may arrive sooner and have a greater impact than previously expected. The rapid pace of investment also reflects growing confidence that the technology is becoming commercially useful rather than remaining a research project. The opportunities and concerns Supporters argue that Agentic AI could make software far more useful. Businesses could automate routine processes. Consumers could spend less time on repetitive digital tasks. Researchers could analyse information more quickly. Developers could accelerate software creation. At the same time, concerns remain. The more autonomy AI systems receive, the more important reliability, security and oversight become. Questions around accountability, data privacy and workplace disruption continue to be debated by governments, researchers and technology companies. The challenge for the industry is finding the right balance between capability and control. A glimpse of the next internet era The rise of Agentic AI may represent a larger shift than many people realise. For decades, software has largely required users to navigate menus, click buttons and manage workflows themselves. AI agents could gradually change that model. Instead of telling software exactly what to do, people may increasingly describe a goal and allow AI systems to determine how to achieve it. Whether that future arrives in months or years remains uncertain. What is becoming increasingly clear, however, is that the conversation in the technology industry has already moved beyond chatbots. The focus is now shifting towards AI systems that can act, adapt and solve problems with greater independence.

The AI industry is rapidly moving towards Agentic AI, a new generation of systems designed to complete tasks rather than simply answer questions. Technology leaders believe AI agents could reshape software, work and productivity faster than many expected.

It's happening faster than expected: How Agentic AI is reshaping the tech industry

Within 48 hours of each other in the first week of June 2026, Nvidia and Google quietly released two open-source AI models that represent fundamentally different visions of what free AI should look like. One is a 550-billion-parameter reasoning beast unveiled on a Taipei stage. The other is a 12-billion-parameter compact model that runs on a regular office laptop. They are not competitors. They are different species and understanding the difference matters.&nbsp;Nvidia unveiled Nemotron 3 Ultra on June 1 at Computex 2026 in Taipei. Google followed with Gemma 4 12B on June 3. The timing felt like a coincidence, but the contrast was too instructive to ignore.&nbsp;Nvidia Nemotron 3 Ultra: The Cloud-Scale Reasoning Machine&nbsp;Nemotron 3 Ultra is a 550-billion-parameter model but that number requires a qualifier. Only 55 billion parameters are active per token. Nvidia built it using a Mixture-of-Experts (MoE) architecture, which means the model activates only the relevant subset of its brain for any given task rather than running the entire network. This design decision is why it can deliver over 300 tokens per second and run 3 to 6 times faster than competing models at similar quality levels.&nbsp;The architecture itself is a hybrid Mamba-Transformer MoE, a combination of state-space models and attention mechanisms that lets the model handle context windows of up to 1 million tokens. For reference, 1 million tokens is roughly the size of a 750-page book that the model can hold in memory and reason over simultaneously. This is not a chatbot. This is infrastructure for complex, long-running AI agents.&nbsp;On the Artificial Analysis Intelligence Index, which measures reasoning ability, programming, general knowledge, and agent performance, Nemotron 3 Ultra scores 48, the highest score ever achieved by an American open-weight model. To put that in perspective, Google's larger Gemma 4 31B scores 39 on the same index. Nvidia's model outperforms it by 12 points while being significantly faster. China's Kimi K2.6 currently sits at 54, but Nemotron closes the gap considerably.&nbsp;Nvidia released the full model weights on Hugging Face for free. The training recipe is also publicly available in the NeMo repository. Enterprises can download and run it in production at no licensing cost. Glean, the enterprise search company, has already integrated Nemotron 3 Ultra into its platform, citing its cost efficiency for complex agentic workflows.&nbsp;Google Gemma 4 12B: The Laptop-Scale Multimodal Pioneer&nbsp;Gemma 4 12B is a completely different kind of animal. At approximately 11.95 billion parameters, it is small enough to run entirely on a standard enterprise laptop using just 16GB of VRAM or unified memory, the kind of memory that comes standard on a modern MacBook Pro with an M-series chip. No cloud subscription. No API calls. No data leaving your machine.&nbsp;What makes Gemma 4 12B technically remarkable is its encoder-free unified architecture. Traditional multimodal AI models bolt on separate processing modules — an image encoder, an audio encoder and stitch their outputs together before feeding them to the language model. Gemma 4 12B eliminates all of that. Raw audio waveforms and visual patches flow directly into the core language model backbone without any secondary processing stage. The result is lower latency, lower memory overhead, and a single model that natively understands text, images, audio, and video.&nbsp;The context window is 256,000 tokens, considerably smaller than Nemotron's 1 million, but entirely appropriate for the kinds of tasks this model is designed for. On mathematical reasoning benchmarks, Gemma 4 12B posted 77.5 percent on AIME 2026 and 79.7 percent on MATH-Vision, numbers that approach Google's own larger 26B MoE model. For a 12B model running locally on consumer hardware, those are genuinely surprising results.&nbsp;Google released Gemma 4 12B under an Apache 2.0 license, one of the most permissive open-source licenses available and made it immediately available on Hugging Face and Kaggle. It integrates with standard deployment frameworks including vLLM, SGLang, MLX, and llama.cpp.&nbsp;Two Models, Two Different Problems They Are Solving&nbsp;The comparison is not really about which model is better. It is about what problem you are trying to solve.&nbsp;Nemotron 3 Ultra is built for developers and enterprises running large-scale agentic AI systems, multi-step workflows, autonomous coding agents, complex document analysis over massive context windows. It requires serious hardware infrastructure to deploy, but once deployed, it delivers near-frontier performance at open-source pricing. This is the model you reach for when you are building AI pipelines that need to reason over thousands of pages of documentation, write and execute code autonomously, or manage complex multi-agent orchestration.&nbsp;Gemma 4 12B is built for a completely different scenario: the edge, the device, the individual. A journalist who wants to run a private transcription and analysis model without sending recordings to a third-party server. A doctor who needs a diagnostic reasoning tool that processes patient data entirely on-premises. A developer in a bandwidth-constrained environment building a multimodal application that needs to work offline. For these use cases, Gemma 4 12B is not a compromise, it is the right choice.&nbsp;Why This Week Matters for Open Source AI&nbsp;The release of both models in the same week is not just news. It marks a maturation point in the open-source AI ecosystem. Until recently, the dominant narrative was that open-source models lagged behind closed frontier models by a generation. That gap has narrowed dramatically.&nbsp;What Nvidia and Google have collectively demonstrated is that open-weight AI can now cover both ends of the performance spectrum, from a 12B model running on consumer hardware to a 550B model matching or exceeding proprietary frontier performance on reasoning benchmarks, without a licensing fee, without an API key, and without dependency on any cloud provider.&nbsp;That is a structural shift. Developers who previously had to choose between capability and cost now have credible open alternatives at both ends. The next wave of AI applications will increasingly be built not on top of closed APIs but on models like these, deployed locally, customised freely, and owned entirely by the teams building them.&nbsp;Two models, released within 48 hours, pointing in opposite directions. Both of them, in their own way, signal where AI is heading next.

Nvidia's Nemotron 3 Ultra and Google's Gemma 4 12B were released within days of each other in June 2026. Here is what makes them fundamentally different and why both matter for the future of open source AI.

Within 48 hours of each other in the first week of June 2026, Nvidia and Google quietly released two open-source AI models that represent fundamentally different visions of what free AI should look like. One is a 550-billion-parameter reasoning beast unveiled on a Taipei stage. The other is a 12-billion-parameter compact model that runs on a regular office laptop. They are not competitors. They are different species and understanding the difference matters. Nvidia unveiled Nemotron 3 Ultra on June 1 at Computex 2026 in Taipei. Google followed with Gemma 4 12B on June 3. The timing felt like a coincidence, but the contrast was too instructive to ignore. Nvidia Nemotron 3 Ultra: The Cloud-Scale Reasoning Machine Nemotron 3 Ultra is a 550-billion-parameter model but that number requires a qualifier. Only 55 billion parameters are active per token. Nvidia built it using a Mixture-of-Experts (MoE) architecture, which means the model activates only the relevant subset of its brain for any given task rather than running the entire network. This design decision is why it can deliver over 300 tokens per second and run 3 to 6 times faster than competing models at similar quality levels. The architecture itself is a hybrid Mamba-Transformer MoE, a combination of state-space models and attention mechanisms that lets the model handle context windows of up to 1 million tokens. For reference, 1 million tokens is roughly the size of a 750-page book that the model can hold in memory and reason over simultaneously. This is not a chatbot. This is infrastructure for complex, long-running AI agents. On the Artificial Analysis Intelligence Index, which measures reasoning ability, programming, general knowledge, and agent performance, Nemotron 3 Ultra scores 48, the highest score ever achieved by an American open-weight model. To put that in perspective, Google's larger Gemma 4 31B scores 39 on the same index. Nvidia's model outperforms it by 12 points while being significantly faster. China's Kimi K2.6 currently sits at 54, but Nemotron closes the gap considerably. Nvidia released the full model weights on Hugging Face for free. The training recipe is also publicly available in the NeMo repository. Enterprises can download and run it in production at no licensing cost. Glean, the enterprise search company, has already integrated Nemotron 3 Ultra into its platform, citing its cost efficiency for complex agentic workflows. Google Gemma 4 12B: The Laptop-Scale Multimodal Pioneer Gemma 4 12B is a completely different kind of animal. At approximately 11.95 billion parameters, it is small enough to run entirely on a standard enterprise laptop using just 16GB of VRAM or unified memory, the kind of memory that comes standard on a modern MacBook Pro with an M-series chip. No cloud subscription. No API calls. No data leaving your machine. What makes Gemma 4 12B technically remarkable is its encoder-free unified architecture. Traditional multimodal AI models bolt on separate processing modules — an image encoder, an audio encoder and stitch their outputs together before feeding them to the language model. Gemma 4 12B eliminates all of that. Raw audio waveforms and visual patches flow directly into the core language model backbone without any secondary processing stage. The result is lower latency, lower memory overhead, and a single model that natively understands text, images, audio, and video. The context window is 256,000 tokens, considerably smaller than Nemotron's 1 million, but entirely appropriate for the kinds of tasks this model is designed for. On mathematical reasoning benchmarks, Gemma 4 12B posted 77.5 percent on AIME 2026 and 79.7 percent on MATH-Vision, numbers that approach Google's own larger 26B MoE model. For a 12B model running locally on consumer hardware, those are genuinely surprising results. Google released Gemma 4 12B under an Apache 2.0 license, one of the most permissive open-source licenses available and made it immediately available on Hugging Face and Kaggle. It integrates with standard deployment frameworks including vLLM, SGLang, MLX, and llama.cpp. Two Models, Two Different Problems They Are Solving The comparison is not really about which model is better. It is about what problem you are trying to solve. Nemotron 3 Ultra is built for developers and enterprises running large-scale agentic AI systems, multi-step workflows, autonomous coding agents, complex document analysis over massive context windows. It requires serious hardware infrastructure to deploy, but once deployed, it delivers near-frontier performance at open-source pricing. This is the model you reach for when you are building AI pipelines that need to reason over thousands of pages of documentation, write and execute code autonomously, or manage complex multi-agent orchestration. Gemma 4 12B is built for a completely different scenario: the edge, the device, the individual. A journalist who wants to run a private transcription and analysis model without sending recordings to a third-party server. A doctor who needs a diagnostic reasoning tool that processes patient data entirely on-premises. A developer in a bandwidth-constrained environment building a multimodal application that needs to work offline. For these use cases, Gemma 4 12B is not a compromise, it is the right choice. Why This Week Matters for Open Source AI The release of both models in the same week is not just news. It marks a maturation point in the open-source AI ecosystem. Until recently, the dominant narrative was that open-source models lagged behind closed frontier models by a generation. That gap has narrowed dramatically. What Nvidia and Google have collectively demonstrated is that open-weight AI can now cover both ends of the performance spectrum, from a 12B model running on consumer hardware to a 550B model matching or exceeding proprietary frontier performance on reasoning benchmarks, without a licensing fee, without an API key, and without dependency on any cloud provider. That is a structural shift. Developers who previously had to choose between capability and cost now have credible open alternatives at both ends. The next wave of AI applications will increasingly be built not on top of closed APIs but on models like these, deployed locally, customised freely, and owned entirely by the teams building them. Two models, released within 48 hours, pointing in opposite directions. Both of them, in their own way, signal where AI is heading next.

Both Nvidia and Google dropped powerful open-source AI models within days of each other in June 2026 — but they solve completely different problems. Here is what makes them two distinct species of AI.

Nvidia's Nemotron 3 Ultra vs Google's Gemma 4 12B: Understanding the difference between two new species of AI

Artificial intelligence (AI) adoption is growing quickly across industries as companies move from testing AI tools to using them in daily operations. But as businesses expand the use of AI agents and automated systems, many are now facing a larger challenge governance.Several enterprises are discovering that building AI tools is only one part of the process. Managing those systems safely, tracking how they work, and ensuring compliance are becoming equally important. According to a report by Boston Consulting Group (BCG), India’s enterprise AI market is expected to grow to nearly $17 billion by 2027. The report says much of this growth will come from companies moving AI projects out of pilot stages and into full-scale deployment.<h2>Why is governance becoming important?</h2>As AI agents begin handling customer support, internal workflows, operations, and data analysis, companies are under immense pressure to ensure these systems remain accountable and transparent. Experts say enterprises now need systems that can monitor AI behaviour, track decisions, manage risks, and maintain compliance with regulations. This is becoming more important as countries introduce stricter AI rules. In Europe, the EU AI Act has increased focus on responsible AI deployment, while businesses globally are facing rising concerns around privacy, security and accountability. Industry analysts believe governance could become one of the most important parts of the enterprise AI market over the next few years.<h2>Google Cloud expands focus on enterprise AI</h2>The growing focus on governance can also be seen in partnerships between cloud companies and AI-focused firms. In April 2026, Google Cloud selected nine “AI-native services partners” as part of its Gemini Enterprise transformation programme, announced at Google Cloud Next 2026 in Las Vegas. The programme is backed by a reported $750 million investment aimed at helping enterprises move from AI experimentation to production-level deployment. Among the selected firms was Covasant Technologies, an AI-native Services-as-Software company with operations across the United States, India, the United Kingdom, and the UAE. Google Cloud said the collaboration would focus on helping enterprises build autonomous AI systems that are integrated into business operations while maintaining oversight and governance. Victor Morales, Vice President of Global System Integrators Partnerships at Google Cloud, said the partnership aims to help customers use Gemini Enterprise to build “autonomous systems” that are closely connected with enterprise operations.<h2>AI-first firms are gaining attention</h2>The rise of AI governance has also brought attention to a new category of firms described as “AI-first” companies. Research firm Constellation Research recently included Covasant Technologies in its inaugural 2026 list of AI-First Service Firms. The list highlighted companies building enterprise solutions using AI-native operating models and digital labour systems. At the same time, enterprises are increasingly looking at certifications and compliance standards before deploying AI systems at scale. Covasant Technologies also achieved the ISO/IEC 42001:2023 accredited certification, which is considered to be one of the first international standards focused on AI management systems. Industry experts say certifications like ISO 42001 may become more important in enterprise procurement discussions as companies seek stronger governance structures for AI deployments.<h2>The next phase of enterprise AI</h2>The enterprise AI market is now entering a new phase where the focus is shifting from experimentation to long-term management. Companies are expected to invest more in governance tools, compliance systems and lifecycle management platforms as AI adoption increases. Analysts say enterprises that successfully balance AI innovation with accountability may gain an advantage as competition in the sector grows.While AI adoption continues to rise globally, experts believe that the companies that can govern AI effectively in addition to deploying it seamlessly could shape the next phase of enterprise AI technology.

Artificial intelligence (AI) adoption is growing quickly across industries as companies move from testing AI tools to using them in daily operations. But as businesses expand the use of AI agents and automated systems, many are now facing a larger challenge governance.

Artificial intelligence (AI) adoption is growing quickly across industries as companies move from testing AI tools to using them in daily operations. But as businesses expand the use of AI agents and automated systems, many are now facing a larger challenge governance. Several enterprises are discovering that building AI tools is only one part of the process. Managing those systems safely, tracking how they work, and ensuring compliance are becoming equally important. According to a report by Boston Consulting Group (BCG), India’s enterprise AI market is expected to grow to nearly $17 billion by 2027. The report says much of this growth will come from companies moving AI projects out of pilot stages and into full-scale deployment. Why is governance becoming important? As AI agents begin handling customer support, internal workflows, operations, and data analysis, companies are under immense pressure to ensure these systems remain accountable and transparent. Experts say enterprises now need systems that can monitor AI behaviour, track decisions, manage risks, and maintain compliance with regulations. This is becoming more important as countries introduce stricter AI rules. In Europe, the EU AI Act has increased focus on responsible AI deployment, while businesses globally are facing rising concerns around privacy, security and accountability. Industry analysts believe governance could become one of the most important parts of the enterprise AI market over the next few years. Google Cloud expands focus on enterprise AI The growing focus on governance can also be seen in partnerships between cloud companies and AI-focused firms. In April 2026, Google Cloud selected nine “AI-native services partners” as part of its Gemini Enterprise transformation programme, announced at Google Cloud Next 2026 in Las Vegas. The programme is backed by a reported $750 million investment aimed at helping enterprises move from AI experimentation to production-level deployment. Among the selected firms was Covasant Technologies, an AI-native Services-as-Software company with operations across the United States, India, the United Kingdom, and the UAE. Google Cloud said the collaboration would focus on helping enterprises build autonomous AI systems that are integrated into business operations while maintaining oversight and governance. Victor Morales, Vice President of Global System Integrators Partnerships at Google Cloud, said the partnership aims to help customers use Gemini Enterprise to build “autonomous systems” that are closely connected with enterprise operations. AI-first firms are gaining attention The rise of AI governance has also brought attention to a new category of firms described as “AI-first” companies. Research firm Constellation Research recently included Covasant Technologies in its inaugural 2026 list of AI-First Service Firms. The list highlighted companies building enterprise solutions using AI-native operating models and digital labour systems. At the same time, enterprises are increasingly looking at certifications and compliance standards before deploying AI systems at scale. Covasant Technologies also achieved the ISO/IEC 42001:2023 accredited certification, which is considered to be one of the first international standards focused on AI management systems. Industry experts say certifications like ISO 42001 may become more important in enterprise procurement discussions as companies seek stronger governance structures for AI deployments. The next phase of enterprise AI The enterprise AI market is now entering a new phase where the focus is shifting from experimentation to long-term management. Companies are expected to invest more in governance tools, compliance systems and lifecycle management platforms as AI adoption increases. Analysts say enterprises that successfully balance AI innovation with accountability may gain an advantage as competition in the sector grows. While AI adoption continues to rise globally, experts believe that the companies that can govern AI effectively in addition to deploying it seamlessly could shape the next phase of enterprise AI technology.

Artificial intelligence (AI) adoption is growing quickly across industries as companies move from testing AI tools to using them in daily operations. But as businesses expand the use of AI agents and automated systems, many are now facing a larger challenge governance.

AI adoption is rising fast: Key things enterprises need before scaling AI agents

Google has announced a major shift in its artificial intelligence strategy, with CEO Sundar Pichai focusing on bringing AI agents into everyday products used by billions of people.Speaking at the Google I/O 2026 developer conference, Pichai said the company is moving from building AI tools for developers and enterprises to delivering agent-based experiences directly to consumers.&nbsp;He noted that Google now has 13 products with over 1 billion users each, including five products used by more than 3 billion people, highlighting the scale at which these AI features can be deployed.<h2>Google Search gets its biggest update in years</h2>One of the key announcements was a major update to Google Search, described as the biggest change in over 25 years.The search box is being redesigned to support longer, more conversational queries. Users will now be able to:Ask complex questionsUpload images, videos or filesGet AI-assisted suggestions while typing&nbsp;Google is combining features like AI Overviews and AI Mode into a single experience. According to company data, AI Overviews already has over 2.5 billion monthly users, while AI Mode has crossed 1 billion monthly users.<h2>Introduction of AI agents in search</h2>Google also introduced Search Agents, which are designed to work in the background and assist users continuously.These agents can:Track market trendsMonitor shopping dealsFollow sports updatesAlert users about relevant changesFor example, users can set up an agent to track stock prices or housing listings based on specific criteria. This marks a shift from manual searching to automated information tracking.<h2>Gemini Spark and personal AI assistants</h2>Google is launching a new personal AI assistant called Gemini Spark within the Gemini app.This always-on assistant can:Summarise emails and meetingsCreate documents and reportsManage tasks and schedules&nbsp;It integrates with tools like Gmail, Docs and Calendar, helping users manage their daily work more efficiently. The system runs on cloud-based infrastructure, allowing it to operate continuously in the background.<h2>AI-powered tools and app creation</h2>Google is also enabling users to create simple applications directly within Search using its AI tools.Users can build:Fitness trackersEvent plannersBudget toolsThese tools can pull real-time data such as weather updates, reviews and location-based information, making them more dynamic and personalised.<h2>Move towards advanced AI models</h2>Alongside product updates, Google introduced a new family of AI models called Gemini Omni. These models can process multiple types of input, including text, images, audio and video.According to Demis Hassabis, the development represents progress towards artificial general intelligence (AGI), where systems can perform a wide range of tasks.The first version, Gemini Omni Flash, will be available across Google platforms, including the Gemini app and YouTube tools.<h2>What this means for users and the industry</h2>Google’s latest announcements show a shift from traditional search and apps to AI-driven experiences that can act on behalf of users.Instead of manually searching and managing tasks, users may increasingly rely on AI agents to handle routine activities.The move also reflects growing competition in the AI space, as companies focus on integrating AI deeply into their products.<h2>What happens next</h2>Many of these features will roll out gradually over the coming months, with some initially available to premium users and testers.As adoption grows, Google’s agent-based approach could change how people interact with digital services, moving from simple queries to continuous, AI-assisted workflows.

Google has announced a major shift in its artificial intelligence strategy, with CEO Sundar Pichai focusing on bringing AI agents into everyday products used by billions of people.


Google has announced a major shift in its artificial intelligence strategy, with CEO Sundar Pichai focusing on bringing AI agents into everyday products used by billions of people. Speaking at the Google I/O 2026 developer conference, Pichai said the company is moving from building AI tools for developers and enterprises to delivering agent-based experiences directly to consumers. He noted that Google now has 13 products with over 1 billion users each, including five products used by more than 3 billion people, highlighting the scale at which these AI features can be deployed. Google Search gets its biggest update in years One of the key announcements was a major update to Google Search, described as the biggest change in over 25 years. The search box is being redesigned to support longer, more conversational queries. Users will now be able to: Ask complex questions Upload images, videos or files Get AI-assisted suggestions while typing Google is combining features like AI Overviews and AI Mode into a single experience. According to company data, AI Overviews already has over 2.5 billion monthly users, while AI Mode has crossed 1 billion monthly users. Introduction of AI agents in search Google also introduced Search Agents, which are designed to work in the background and assist users continuously. These agents can: Track market trends Monitor shopping deals Follow sports updates Alert users about relevant changes For example, users can set up an agent to track stock prices or housing listings based on specific criteria. This marks a shift from manual searching to automated information tracking. Gemini Spark and personal AI assistants Google is launching a new personal AI assistant called Gemini Spark within the Gemini app. This always-on assistant can: Summarise emails and meetings Create documents and reports Manage tasks and schedules It integrates with tools like Gmail, Docs and Calendar, helping users manage their daily work more efficiently. The system runs on cloud-based infrastructure, allowing it to operate continuously in the background. AI-powered tools and app creation Google is also enabling users to create simple applications directly within Search using its AI tools. Users can build: Fitness trackers Event planners Budget tools These tools can pull real-time data such as weather updates, reviews and location-based information, making them more dynamic and personalised. Move towards advanced AI models Alongside product updates, Google introduced a new family of AI models called Gemini Omni. These models can process multiple types of input, including text, images, audio and video. According to Demis Hassabis, the development represents progress towards artificial general intelligence (AGI), where systems can perform a wide range of tasks. The first version, Gemini Omni Flash, will be available across Google platforms, including the Gemini app and YouTube tools. What this means for users and the industry Google’s latest announcements show a shift from traditional search and apps to AI-driven experiences that can act on behalf of users. Instead of manually searching and managing tasks, users may increasingly rely on AI agents to handle routine activities. The move also reflects growing competition in the AI space, as companies focus on integrating AI deeply into their products. What happens next Many of these features will roll out gradually over the coming months, with some initially available to premium users and testers. As adoption grows, Google’s agent-based approach could change how people interact with digital services, moving from simple queries to continuous, AI-assisted workflows.

Google has announced a major shift in its artificial intelligence strategy, with CEO Sundar Pichai focusing on bringing AI agents into everyday products used by billions of people.

AI in everyday life: How Google plans to transform user experience

The way people use smartphones may change completely in the coming years. According to Carl Pei, CEO of Nothing, traditional mobile apps could disappear as artificial intelligence becomes more advanced.Speaking at the SXSW conference in Austin, Pei said the current app-based system is outdated and may soon be replaced by AI agents that can act on behalf of users.<h2>Why Carl Pei says apps may disappear</h2>Pei believes that today’s smartphone experience has not changed much in the past two decades.Users still rely on:Lock screens and home screensMultiple apps for different tasksApp stores to install softwareHe said this model creates unnecessary steps for simple tasks. For example, planning to meet a friend for coffee may involve messaging apps, maps, ride-hailing services and calendars.“It’s very hard to get things done on a phone,” Pei said, pointing out that users often switch between several apps to complete one task.According to him, AI agents could remove this complexity by handling tasks directly.

<h2>What is an AI-first smartphone?</h2>Pei described a future device powered by AI agents instead of apps.In this system:Users express an intention (for example, “book a trip” or “meet a friend”)The AI understands the requestThe system completes the task automaticallyInstead of opening apps and navigating menus, users would rely on AI to manage everything in the background. Pei said the goal is to build devices that “know you very well” and can act without constant input.<h2>How AI agents could work</h2>The transition to AI-first devices may happen in stages.Pei explained that the first stage already visible today is AI tools that can perform tasks like booking flights or ordering food.However, he called this stage “basic”.The next stage, according to him, is more advanced. AI systems will begin to understand long-term user behaviour and preferences.For example:Suggesting healthier habitsRecommending activities based on past behaviour Managing schedules automaticallyHe compared this to features like memory in AI tools such as ChatGPT, where systems learn from past interactions.<h2>Why current smartphones may change</h2>Pei argued that modern technology has advanced significantly, but smartphone design has not kept up.He said the current system still relies on humans manually interacting with apps. In contrast, future systems may be designed for AI agents to operate directly.“The future is not the agent using a human interface,” he said. “You need to create an interface for the agent.”This means smartphones may shift from app-based interfaces to AI-driven systems working in the background.<h2>Will apps disappear soon?</h2>Pei clarified that apps will not vanish immediately. Current systems, including Nothing’s own software, still rely on apps. However, over time, AI agents may take over more tasks, reducing the need for users to interact with apps directly.The idea reflects a broader trend in the tech industry, where companies are building AI agents that can perform actions across platforms instead of simply providing information.<h2>The bigger picture</h2>Pei’s comments come as AI companies and device makers explore new ways to integrate artificial intelligence into everyday technology.The shift from apps to AI agents could change how people use smartphones, manage tasks and interact with digital services.While the transition may take time, the direction is becoming clearer: future devices may focus less on screens and apps, and more on intelligent systems that act on behalf of users.&nbsp;

At the SXSW, Carl Pei said traditional smartphone apps could eventually disappear as AI agents take over everyday tasks. He argued that the current app-based system is outdated and inefficient, requiring users to switch between multiple apps for simple actions. 

The way people use smartphones may change completely in the coming years. According to Carl Pei, CEO of Nothing, traditional mobile apps could disappear as artificial intelligence becomes more advanced. Speaking at the SXSW conference in Austin, Pei said the current app-based system is outdated and may soon be replaced by AI agents that can act on behalf of users. Why Carl Pei says apps may disappear Pei believes that today’s smartphone experience has not changed much in the past two decades. Users still rely on: Lock screens and home screens Multiple apps for different tasks App stores to install software He said this model creates unnecessary steps for simple tasks. For example, planning to meet a friend for coffee may involve messaging apps, maps, ride-hailing services and calendars. “It’s very hard to get things done on a phone,” Pei said, pointing out that users often switch between several apps to complete one task. According to him, AI agents could remove this complexity by handling tasks directly.

At the SXSW, Carl Pei said traditional smartphone apps could eventually disappear as AI agents take over everyday tasks. He argued that the current app-based system is outdated and inefficient, requiring users to switch between multiple apps for simple actions.&nbsp;

‘Apps may disappear in the future’: AI agents will take over, says Nothing CEO Carl Pei

OpenAI has launched Frontier, a new enterprise platform designed to help organisations build and manage artificial intelligence agents that can work across teams and systems.According to OpenAI, Frontier is aimed at companies that want to move beyond using AI as a simple chatbot or productivity tool, and instead use it as a shared digital co-worker across the organisation.The platform is currently being rolled out to a limited number of organisations, with a wider release expected later.&nbsp;<h2>What is OpenAI’s Frontier?</h2>It is an enterprise workspace where multiple AI agents can be created, trained and deployed together. These agents can operate in the same context as human employees, using company data, tools and rules.OpenAI says Frontier is designed for “real work”, not one-off prompts. Instead of asking an AI tool a single question, companies can set up agents that handle tasks across departments, systems and workflows.According to OpenAI, early users include large organisations such as Thermo Fisher Scientific, Intuit and Oracle.<h2>How does Frontier work inside a company?</h2>Frontier is built in a way that mirrors how human employees are onboarded. AI agents are given clear instructions, access limits and training before they start working.The platform allows organisations to connect:&nbsp;Internal data such as CRMs and data warehouses&nbsp;Identity rules, permissions and restrictions&nbsp;Clear definitions of what good and poor work looks like&nbsp;Access to internal and third-party applications as tools&nbsp;This setup allows AI agents to work across systems, receive feedback and improve over time. OpenAI says this reduces errors and helps agents complete tasks faster.

<blockquote class="twitter-tweet">Introducing OpenAI Frontier—a new platform that helps enterprises build, deploy, and manage AI coworkers that can do real work. <a href="https://t.co/4W0adQzSZ1">https://t.co/4W0adQzSZ1</a>&mdash; OpenAI (@OpenAI) <a href="https://twitter.com/OpenAI/status/2019413712772411528?ref_src=twsrc%5Etfw">February 5, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

<h2>Why Frontier matters for everyday work</h2>Until now, most AI tools have helped individuals work faster at a task level. Frontier shifts AI into a shared role across teams.This means AI agents can support workflows across departments, reduce delays between teams and keep work moving even when people are unavailable. For example, if a team member is on leave, an AI agent can still provide access to information or updates.Experts say this could change how organisations think about productivity, handovers and even work-life balance.&nbsp;<h2>Which jobs and teams will be affected most?</h2>Frontier is not limited to engineers or IT teams. OpenAI says non-technical teams can also use the platform.Roles likely to be affected include project management, customer operations, procurement, marketing, sales and other office functions. Some roles may benefit from support and automation, while others, especially junior or assistant roles, may see tasks reduced or changed.<h2>What leaders and managers should watch out for</h2>Some experts have warned that while Frontier makes AI adoption easier, companies should not rely on it blindly.Jason Schloetzer, an associate professor at Georgetown University, said organisations should still build internal skills and oversight. He warned that companies risk becoming too dependent on a single platform if they do not develop their own understanding of AI systems.He suggested that leaders should train managers to supervise AI agents, set clear rules, and keep humans involved in quality checks and important decisions.<h2>What comes next for Frontier and enterprise AI</h2>Frontier reflects a broader shift in how AI is used at work. Instead of being a standalone tool, AI is becoming part of core organisational infrastructure.OpenAI’s platform is unlikely to be the last of its kind. Other big technology companies are expected to launch similar systems as competition in enterprise AI grows.For businesses, experts say the key will be to balance automation with human judgement, and to prepare teams for working alongside AI agents rather than being replaced by them.

OpenAI has launched Frontier, an enterprise platform that lets organisations deploy AI agents as shared digital co-workers. Frontier allows AI to work across teams, systems and data, shifting AI from chatbots to core workplace infrastructure.

OpenAI has launched Frontier, a new enterprise platform designed to help organisations build and manage artificial intelligence agents that can work across teams and systems. According to OpenAI, Frontier is aimed at companies that want to move beyond using AI as a simple chatbot or productivity tool, and instead use it as a shared digital co-worker across the organisation. The platform is currently being rolled out to a limited number of organisations, with a wider release expected later. What is OpenAI’s Frontier? It is an enterprise workspace where multiple AI agents can be created, trained and deployed together. These agents can operate in the same context as human employees, using company data, tools and rules. OpenAI says Frontier is designed for “real work”, not one-off prompts. Instead of asking an AI tool a single question, companies can set up agents that handle tasks across departments, systems and workflows. According to OpenAI, early users include large organisations such as Thermo Fisher Scientific, Intuit and Oracle. How does Frontier work inside a company? Frontier is built in a way that mirrors how human employees are onboarded. AI agents are given clear instructions, access limits and training before they start working. The platform allows organisations to connect: Internal data such as CRMs and data warehouses Identity rules, permissions and restrictions Clear definitions of what good and poor work looks like Access to internal and third-party applications as tools This setup allows AI agents to work across systems, receive feedback and improve over time. OpenAI says this reduces errors and helps agents complete tasks faster.

OpenAI has launched Frontier, an enterprise platform that lets organisations deploy AI agents as shared digital co-workers. Frontier allows AI to work across teams, systems and data, shifting AI from chatbots to core workplace infrastructure.

What is OpenAI’s Frontier and how it works: All you need to know

Anthropic trains AI agents to audit powerful models for safetyAs artificial intelligence continues to grow more advanced, so do concerns about its safety. Anthropic, which is the AI research company behind Claude, has developed a new solution using AI to audit AI. The company has created a group of digital agents designed to investigate and fix hidden problems in large AI models before they can cause harm.&nbsp;These AI agents are now being used to test Anthropic’s own models, including Opus 4, and are seen as part of a wider move to build trust and accountability in AI development.&nbsp;<h2>Anthropic’s safety system involves three different AI agents, each with a unique task.</h2>The Investigator Agent looks deep into the model’s behaviour, acting like a digital detective. It checks the model’s training and logic to find the root cause of any suspicious patterns. For example, it can look inside the model’s neural network to understand how it’s thinking.&nbsp;Next is the Evaluation Agent, which runs controlled tests on specific issues. If a model responds in a concerning way, this agent tests how often it behaves like that, gathering clear data.&nbsp;Finally, the Red-Teaming Agent interacts with the model in many different ways, trying to trigger unexpected or dangerous responses. Any worrying behaviour it finds is flagged for review by human researchers.&nbsp;Together, these models works like a safety net scanning, testing, and reporting on how these powerful AI systems behave.&nbsp;<h2>What the tests showed</h2>To test the system, Anthropic trained a model with a hidden flaw. It had been taught to act like a sycophant, agreeing with false information to gain approval. The Investigator Agent only found this hidden goal in 13 per cent of cases when working alone.&nbsp;But when ten Investigator Agents worked together, the success rate jumped to 42 per cent. This suggests that, like in human teams, collaboration improves outcomes.&nbsp;The Evaluation Agent spotted problems in 88 per cent of test cases, while the Red-Teaming Agent uncovered 7 out of 10 intentionally hidden issues. This shows strong promise, even though the agents are not yet perfect.&nbsp;<h2>From lab to real-world models</h2>Anthropic has already added these safety agents to its live AI models. They have helped detect risks such as “prefill attacks”, where users start a sentence in a way that leads the AI into generating harmful responses.&nbsp;One of the most serious findings came from the Investigator Agent. It discovered a neural pathway inside the Opus 4 model that was linked to spreading misinformation. When this part was activated, the model ignored safety filters and created a fake news article claiming vaccines cause autism, a false claim that has been widely debunked.This finding highlighted a key issue: tools built to protect us could be misused if not properly controlled.&nbsp;<h2>A shift in AI safety work</h2>The Anthropic team admits these agents are still learning. They can miss subtle issues or get stuck on a single idea. However, they reduce the burden on human teams and allow experts to focus on higher-level decisions.&nbsp;Instead of reviewing each model manually, human researchers now design the agents and guide their efforts. This shift could help scale safety efforts as AI systems grow more powerful and complex.&nbsp;Anthropic’s approach shows how we might use AI to police AI using the same power that creates risks to help reduce them.

Anthropic using AI agents to audit its AI models like Opus 4 for hidden flaws such as misinformation bias and malicious responses. With Investigator, Evaluation and Red-Teaming agents collaborating, the system catches risks earlier. Read more below.

Anthropic trains AI agents to audit powerful models for safety As artificial intelligence continues to grow more advanced, so do concerns about its safety. Anthropic, which is the AI research company behind Claude, has developed a new solution using AI to audit AI. The company has created a group of digital agents designed to investigate and fix hidden problems in large AI models before they can cause harm. These AI agents are now being used to test Anthropic’s own models, including Opus 4, and are seen as part of a wider move to build trust and accountability in AI development. Anthropic’s safety system involves three different AI agents, each with a unique task. The Investigator Agent looks deep into the model’s behaviour, acting like a digital detective. It checks the model’s training and logic to find the root cause of any suspicious patterns. For example, it can look inside the model’s neural network to understand how it’s thinking. Next is the Evaluation Agent, which runs controlled tests on specific issues. If a model responds in a concerning way, this agent tests how often it behaves like that, gathering clear data. Finally, the Red-Teaming Agent interacts with the model in many different ways, trying to trigger unexpected or dangerous responses. Any worrying behaviour it finds is flagged for review by human researchers. Together, these models works like a safety net scanning, testing, and reporting on how these powerful AI systems behave. What the tests showed To test the system, Anthropic trained a model with a hidden flaw. It had been taught to act like a sycophant, agreeing with false information to gain approval. The Investigator Agent only found this hidden goal in 13 per cent of cases when working alone. But when ten Investigator Agents worked together, the success rate jumped to 42 per cent. This suggests that, like in human teams, collaboration improves outcomes. The Evaluation Agent spotted problems in 88 per cent of test cases, while the Red-Teaming Agent uncovered 7 out of 10 intentionally hidden issues. This shows strong promise, even though the agents are not yet perfect. From lab to real-world models Anthropic has already added these safety agents to its live AI models. They have helped detect risks such as “prefill attacks”, where users start a sentence in a way that leads the AI into generating harmful responses. One of the most serious findings came from the Investigator Agent. It discovered a neural pathway inside the Opus 4 model that was linked to spreading misinformation. When this part was activated, the model ignored safety filters and created a fake news article claiming vaccines cause autism, a false claim that has been widely debunked. This finding highlighted a key issue: tools built to protect us could be misused if not properly controlled. A shift in AI safety work The Anthropic team admits these agents are still learning. They can miss subtle issues or get stuck on a single idea. However, they reduce the burden on human teams and allow experts to focus on higher-level decisions. Instead of reviewing each model manually, human researchers now design the agents and guide their efforts. This shift could help scale safety efforts as AI systems grow more powerful and complex. Anthropic’s approach shows how we might use AI to police AI using the same power that creates risks to help reduce them.

Anthropic using AI agents to audit its AI models like Opus 4 for hidden flaws such as misinformation bias and malicious responses. With Investigator, Evaluation and Red-Teaming agents collaborating, the system catches risks earlier. Read more below.

Anthropic using AI-powered agents to detect hidden flaws in AI models

ai-agents

Ai Agents, News on Ai Agents, stories on Ai Agents, articles on Ai Agents, Ai Agents stories at Wion, Ai Agents news, Ai Agents latest news

ai agents

AI gone rogue? CERT-In issues ‘critical’ alert after agents access confidential data at Indian firms

Anthropic finds AI agents bypassing restrictions, moves to limit internet access

AI testing agents: How AI is being used to fix software testing problems

What are OpenAI Dots and how do these always-on AI agents work?

AI agents can work for you, but should they have access to everything?

OpenAI AI Agents hacked 4 websites before Hugging Face, researchers reveal

Meta’s AI gamble: Why Zuckerberg planned teams of just three to five people

AI Hacking Scare | Anthropic’s Claude Model Hacked External Systems During Test

WhatsApp is testing a major AI feature that could change how you use chats

'One chip, trillions of calculations': How AI processors work

What is Prentis? AI lab reportedly seeking $100 million after launching in April

US debates AI kill switch | Lawmakers fear Artificial Intelligence could escape human control

OpenClaw's AI assistant is now on your phone. Here's what changes

From cinema to care: Filmmaker Medha Jaishankar champions AI health tech for marginalised women

TikTok could cut another 300 jobs in Dublin. Here's what we know

It's happening faster than expected: How Agentic AI is reshaping tech industry

Nvidia's Nemotron 3 Ultra vs Google's Gemma 4 12B

AI adoption is rising fast: Key things enterprises need before scaling AI agents

AI in everyday life: How Google plans to transform user experience

Nothing CEO Carl Pei says AI agents could end the app era on smartphones

What is OpenAI’s Frontier and how it works: All you need to know

Anthropic using AI-powered agents to detect hidden flaws in AI models