Anthropic has revealed details about a new experimental AI model called Claude Mythos, which the company says is its most advanced system so far. According to Anthropic, the model has demonstrated a strong ability to detect serious cybersecurity flaws in widely used software systems. Because of its capabilities, the company is not releasing the model publicly yet and is instead working with technology partners to fix the vulnerabilities it discovered.The announcement reflects how AI tools are increasingly being used in cybersecurity research to identify weaknesses in complex software systems.<h2>Claude Mythos identified thousands of software vulnerabilities</h2>In a blog post, Anthropic said Claude Mythos Preview was able to identify thousands of previously unknown vulnerabilities in widely used software.These include “zero-day vulnerabilities”, which are security flaws that software developers have not yet discovered. Zero-day vulnerabilities are particularly concerning because attackers can exploit them before developers have time to fix them.<h2>Anthropic said the AI model identified vulnerabilities across:</h2>Web browsersOperating systemsCore system softwareThe company added that some of the vulnerabilities could potentially allow attackers to escalate access from a normal user account to full control of a system.Vulnerabilities discovered in OpenBSD and Linux systemsAmong the findings was a 27-year-old vulnerability in OpenBSD, an operating system widely used in security-focused infrastructure such as firewalls and servers. The model also identified several vulnerabilities in the Linux kernel, which powers the majority of servers worldwide and forms the backbone of many internet services.Security researchers say the Linux kernel is one of the most widely analysed pieces of software in the world, which makes the discovery of new vulnerabilities particularly notable.According to Anthropic, these flaws could potentially allow attackers to gain elevated access to systems if exploited.<h2>Anthropic working with 40 technology companies</h2>Because of the security implications, Anthropic said it is collaborating with more than 40 technology companies to address the vulnerabilities identified by Claude Mythos.The goal is to fix the security issues before the model is released more broadly. This approach reflects a growing concern in the AI industry: extremely capable AI systems could potentially identify security weaknesses faster than humans, which creates both benefits and risks.Anthropic said controlled testing with industry partners will help ensure that vulnerabilities are addressed responsibly.<h2>Why the model is not yet publicly released</h2>Although Claude Mythos reportedly outperforms earlier models such as Claude Opus 4.6 and other competing AI systems on several benchmarks, Anthropic is taking a cautious approach.The company said the model’s cybersecurity capabilities mean it needs additional testing and safeguards before public deployment.Anthropic CEO Dario Amodei has previously warned that highly capable AI models may present new risks if released without careful evaluation. Earlier reports also suggested that the model’s existence first became known after references appeared in a data leak related to Anthropic’s internal systems.<h2>AI and the future of cybersecurity</h2>Cybersecurity experts increasingly see artificial intelligence as a powerful tool for analysing large codebases and identifying weaknesses that human researchers may overlook.Traditional vulnerability discovery often requires months of manual review by security teams. AI systems can scan code at scale and highlight suspicious patterns within minutes.However, experts say this capability must be handled carefully. The same technology that identifies vulnerabilities could potentially be misused if released without safeguards. Anthropic’s decision to work with industry partners before releasing Claude Mythos suggests that AI companies are becoming more cautious about how powerful models are deployed. As AI research continues to advance, tools like Claude Mythos may play an important role in strengthening digital security across software systems used worldwide.&nbsp;

Anthropic has revealed an experimental AI system called Claude Mythos, which reportedly discovered thousands of previously unknown cybersecurity vulnerabilities, including zero-day flaws in systems such as OpenBSD and the Linux kernel. 

Anthropic has revealed details about a new experimental AI model called Claude Mythos, which the company says is its most advanced system so far. According to Anthropic, the model has demonstrated a strong ability to detect serious cybersecurity flaws in widely used software systems. Because of its capabilities, the company is not releasing the model publicly yet and is instead working with technology partners to fix the vulnerabilities it discovered. The announcement reflects how AI tools are increasingly being used in cybersecurity research to identify weaknesses in complex software systems. Claude Mythos identified thousands of software vulnerabilities In a blog post, Anthropic said Claude Mythos Preview was able to identify thousands of previously unknown vulnerabilities in widely used software. These include “zero-day vulnerabilities”, which are security flaws that software developers have not yet discovered. Zero-day vulnerabilities are particularly concerning because attackers can exploit them before developers have time to fix them. Anthropic said the AI model identified vulnerabilities across: Web browsers Operating systems Core system software The company added that some of the vulnerabilities could potentially allow attackers to escalate access from a normal user account to full control of a system. Vulnerabilities discovered in OpenBSD and Linux systems Among the findings was a 27-year-old vulnerability in OpenBSD, an operating system widely used in security-focused infrastructure such as firewalls and servers. The model also identified several vulnerabilities in the Linux kernel, which powers the majority of servers worldwide and forms the backbone of many internet services. Security researchers say the Linux kernel is one of the most widely analysed pieces of software in the world, which makes the discovery of new vulnerabilities particularly notable. According to Anthropic, these flaws could potentially allow attackers to gain elevated access to systems if exploited. Anthropic working with 40 technology companies Because of the security implications, Anthropic said it is collaborating with more than 40 technology companies to address the vulnerabilities identified by Claude Mythos. The goal is to fix the security issues before the model is released more broadly. This approach reflects a growing concern in the AI industry: extremely capable AI systems could potentially identify security weaknesses faster than humans, which creates both benefits and risks. Anthropic said controlled testing with industry partners will help ensure that vulnerabilities are addressed responsibly. Why the model is not yet publicly released Although Claude Mythos reportedly outperforms earlier models such as Claude Opus 4.6 and other competing AI systems on several benchmarks, Anthropic is taking a cautious approach. The company said the model’s cybersecurity capabilities mean it needs additional testing and safeguards before public deployment. Anthropic CEO Dario Amodei has previously warned that highly capable AI models may present new risks if released without careful evaluation. Earlier reports also suggested that the model’s existence first became known after references appeared in a data leak related to Anthropic’s internal systems. AI and the future of cybersecurity Cybersecurity experts increasingly see artificial intelligence as a powerful tool for analysing large codebases and identifying weaknesses that human researchers may overlook. Traditional vulnerability discovery often requires months of manual review by security teams. AI systems can scan code at scale and highlight suspicious patterns within minutes. However, experts say this capability must be handled carefully. The same technology that identifies vulnerabilities could potentially be misused if released without safeguards. Anthropic’s decision to work with industry partners before releasing Claude Mythos suggests that AI companies are becoming more cautious about how powerful models are deployed. As AI research continues to advance, tools like Claude Mythos may play an important role in strengthening digital security across software systems used worldwide.

Anthropic has revealed an experimental AI system called Claude Mythos, which reportedly discovered thousands of previously unknown cybersecurity vulnerabilities, including zero-day flaws in systems such as OpenBSD and the Linux kernel.&nbsp;

‘Claude Mythos in cybersecurity’: How Anthropic’s new AI model can find hidden software vulnerabilities

AI agents: promise vs realityAgentic AI – Intelligent software which is designed to take action on its own without human intervention is gaining momentum in tech. From automation of tasks to making routine decisions, AI agents are being explored across most of the industries. But according to a recent&nbsp;<a href="https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027">Gartner report</a>, the excitement may be running ahead of results.&nbsp;The report predicts that around 40 per cent of agentic AI projects will be cancelled by 2027, mainly due to rising costs, vague business outcomes, and the lack of mature risk and governance frameworks.<h2>What is agentic AI, and why is it trending?</h2>Agentic AI is not like traditional automation tools. These systems can make decisions, initiate tasks, and adjust over time based on inputs. Companies are experimenting with these agents to improve productivity, reduce human error, and scale operations.&nbsp;However, according to Anushree Verma, Senior Director Analyst at Gartner, many current projects are driven by hype rather than a clear understanding of business need. “Most agentic AI projects as of now are early-stage experiments that are mostly driven by hype and are often misapplied,” she said.<h2>Why are so many projects failing?</h2>Despite growing interest, implementation is proving challenging. Gartner’s 2025 survey of over 3,400 professionals found that only 19 per cent had made significant investments in agentic AI. A further 42 per cent were investing cautiously, while 31 per cent remained undecided.<h2>Issues which are holding back progress: &nbsp;</h2>High costs: Beyond development, companies face expenses related to infrastructure, compliance, and training.&nbsp;Legacy systems: Many traditional businesses struggle to integrate agentic AI into older platforms without major redesign.&nbsp;Lack of clear ROI:&nbsp; No measurable outcomes; many projects lose momentum.&nbsp;Misuse of the term: Many companies are relabelling chatbots or process automation tools as “AI agents”. &nbsp;Gartner reports highlight that only around 130 providers globally offer agentic solutions that truly meet the criteria of autonomy.<h2>Where agentic AI could succeed</h2>Despite current setbacks, reports also say it remains optimistic about the long-term impact. By 2028, it forecasts that 15 per cent of routine business decisions will be made by AI agents, and a third of enterprise software will include some level of agentic functionality.&nbsp;Areas where agentic AI is already showing potential include:&nbsp;Decision-making assisting in fraud detection or supply chain alerts.&nbsp;Workflow automation – managing complex claims or approvals.&nbsp;Productivity tools – automating IT or HR support at scale.&nbsp;Agentic AI is not a passing trend, but success depends on clear goals, strong governance, and a willingness to redesign how work gets done. While many projects may fail in the near term, those that align the technology with real-world problems stand a better chance of delivering long-term value.

Agentic AI, software that acts without human input, is trending but often fails due to high costs, unclear goals, and poor integration. Gartners report says 40 per cent of projects may be cancelled by 2027, but long term success is possible with clear use cases and governance.

AI agents: promise vs reality Agentic AI – Intelligent software which is designed to take action on its own without human intervention is gaining momentum in tech. From automation of tasks to making routine decisions, AI agents are being explored across most of the industries. But according to a recent Gartner report, the excitement may be running ahead of results. The report predicts that around 40 per cent of agentic AI projects will be cancelled by 2027, mainly due to rising costs, vague business outcomes, and the lack of mature risk and governance frameworks. What is agentic AI, and why is it trending? Agentic AI is not like traditional automation tools. These systems can make decisions, initiate tasks, and adjust over time based on inputs. Companies are experimenting with these agents to improve productivity, reduce human error, and scale operations. However, according to Anushree Verma, Senior Director Analyst at Gartner, many current projects are driven by hype rather than a clear understanding of business need. “Most agentic AI projects as of now are early-stage experiments that are mostly driven by hype and are often misapplied,” she said. Why are so many projects failing? Despite growing interest, implementation is proving challenging. Gartner’s 2025 survey of over 3,400 professionals found that only 19 per cent had made significant investments in agentic AI. A further 42 per cent were investing cautiously, while 31 per cent remained undecided. Issues which are holding back progress: High costs: Beyond development, companies face expenses related to infrastructure, compliance, and training. Legacy systems: Many traditional businesses struggle to integrate agentic AI into older platforms without major redesign. Lack of clear ROI: No measurable outcomes; many projects lose momentum. Misuse of the term: Many companies are relabelling chatbots or process automation tools as “AI agents”. Gartner reports highlight that only around 130 providers globally offer agentic solutions that truly meet the criteria of autonomy. Where agentic AI could succeed Despite current setbacks, reports also say it remains optimistic about the long-term impact. By 2028, it forecasts that 15 per cent of routine business decisions will be made by AI agents, and a third of enterprise software will include some level of agentic functionality. Areas where agentic AI is already showing potential include: Decision-making assisting in fraud detection or supply chain alerts. Workflow automation – managing complex claims or approvals. Productivity tools – automating IT or HR support at scale. Agentic AI is not a passing trend, but success depends on clear goals, strong governance, and a willingness to redesign how work gets done. While many projects may fail in the near term, those that align the technology with real-world problems stand a better chance of delivering long-term value.

Agentic AI, software that acts without human input, is trending but often fails due to high costs, unclear goals, and poor integration. Gartner's report says 40 per cent of projects may be cancelled by 2027, but long-term success is possible with clear use cases and governance.

Future of agentic AI: Why most AI agent projects aren't ready for real-world use

intelligent-software

Intelligent Software, News on Intelligent Software, stories on Intelligent Software, articles on Intelligent Software, Intelligent Software stories at Wion, Intelligent Software news, Intelligent Software latest news

intelligent software

‘Claude Mythos in cybersecurity’: How Anthropic’s new AI model can find hidden software vulnerabilities

Future of agentic AI: Why most AI agent projects aren't ready for real-world use