&imwidth=800&imheight=600&format=webp&quality=medium "After OpenAI's 'scary' agent, Anthropic's Claude goes rogue; AI gains unauthorised 'real-world' access during testing")
AI safety concerns have intensified after Anthropic disclosed that several versions of its Claude model gained unauthorised access to real-world computer systems during security testing. This comes just days after Anthropic rival OpenAI revealed a similar containment failure involving one of its own advanced models.
In a blog post published on Thursday (Jul 30), Anthropic said it reviewed more than 141,000 evaluation runs and found that three different versions of Claude had improperly accessed the systems of three external organisations.
Also read | 'Scary' Rogue AI on the loose? OpenAI agent hit another tech firm's customer, suggests report
How could the AI get 'real-world access'?
Anthropic said the incidents occurred because the AI models unexpectedly had internet access during testing due to a "misunderstanding" between Anthropic and its evaluation partner, Irregular.
Once connected, Claude used "basic techniques, such as exploiting weak passwords and unauthenticated endpoints" to enter external systems, according to the company.
Trending Stories
)
Student denied admission to IIT Kanpur's cybersecurity program hacks websites to prove a point
)
WATCH: 'Absolute chaos' at Spain-Morocco border; thousands breach Ceuta, at least 9 killed
)
Russia-Ukraine war: Putin mobilises more troops for Ukraine's war decisive invasion
)
Bizarre new material discovered in Hiroshima nuclear bombing debris | Gravitas
)
Anthropic reveals India is Claude AI's No. 2 market. But what are Indians actually using the AI for? | WION Decodes
)
This Ramayana actor took ZERO fee from Rs 4000 crore film. Guess who?
)
Indian forces used pellet guns on Jul 20? SC hearing, police diary entry, MLC report and injured protesters' claims
)
The Royals enter the Doctor Who universe: King Charles and Queen Camilla beam into Commonwealth Games in TARDIS
)
23ft python swallows man returning from party in Indonesia; family find body whole inside reptile
&im=FitAndFill=(700,400))
New Friday OTT releases (July 31, 2026): The Vacation Principle, Rao Bahadur, Chum- Watch 6 new movies and TV shows on Netflix, Prime Video and more
)
'There needs to be a declaration of war': Trump wins Senate vote, but Cassidy draws red line on Iran
)
3.6mn people track 24hr Qantas flight from Melbourne to Toulouse: Pilot shifts, fuel tank and other details
)
ED raids uncover ₹600 crore assets in Jharkhand illegal coal mining crackdown
)
582 tonnes, 150 million°C, 60 years: WION Decodes the mind-boggling numbers behind China's 'Artificial Sun'
)
China builds 500+ spy satellites as US space lead shrinks | Gravitas
)
Karan Johar's The Traitors season 2 contestants: Rhea Chakraborty, Mallika Sherawat and Munawar Faruqui in the first batch
)
'Not sure': Trump pours cold water on Ukraine's Patriot missile dream after Zelensky's claim
)
Ex-French pilot who praised Chinese fighters lands in China spy scandal. This is what we know
)
NEET Protests riddled with misinformation? Delhi Police blocks 480 Pakistan-based handles spreading 'fake news' since Operation Sindoor
)
Trump to charge foreign students $100,000 for a shot at a US job? Here's why Indian students should be worried
)
India avoids higher US tariff, secures lower 10% rate. Here's how
)
British Airways flight comes close to 'nosediving' over London
)
Iran claims attack on US military bases in Jordan, Kuwait, alleges F-35s destroyed
)
Nearly Rs 29 crore cash, 15 kg gold found at ex-bus driver's house in West Bengal
)
Kiara Advani reveals why she chose Toxic, says character offered her 'all the meat an actor longs for'
)
'We have special forces': Netanyahu warns against any ICC arrest attempt abroad
)
Hosts India sweep Malaysia for double gold in men’s and women’s Commonwealth Table Tennis finals
)
Lift accident? Maintenance, builders & makers are ALL liable now, rules Supreme Court: 3 major regulations | WION Decodes
)
CJP founder Abhijeet Dipke returns to Maharashtra, receives grand welcome in hometown
)
Russia-Ukraine War: Zelensky fears 500,000 new Russian troops
)
'Hamas will disarm': Trump unveils 'historic' 20-point plan to end Gaza war. What's inside it?
)
Saudi leads 43 nations for new Red Sea naval coalition amid Houthi blockade
)
Delhi Consumer Court fines Indian Railways Rs 25,000 for making RAC passengers share one blanket
)
Brewery offering 'free beer' the day Trump dies has license revoked - 'Will have to pry me out of my taprooms'
)
Trump says Iranian school bombing photos could be AI-generated - 'I don’t think...'
One of the affected models was Mythos 5, Anthropic's most advanced AI system, which is currently available only to a limited group of approved partners.
Anthropic said it is working with Irregular to investigate the incidents and has contacted, or attempted to contact, all three organisations whose systems were accessed.
OpenAI's rogue agent
The disclosure comes less than a week after OpenAI acknowledged that one of its frontier AI models escaped its testing environment, connected to the internet and infiltrated Hugging Face, a platform widely used by developers to host and share AI models and code.
OpenAI later revealed it had identified three additional incidents involving unauthorised access during testing.
The company has since paused parts of its evaluation programme while strengthening its sandboxing systems, which are designed to isolate experimental AI models from external networks.
OpenAI CEO Sam Altman was also expected to meet White House officials on Thursday to discuss the company's upcoming AI models and voluntary government cybersecurity testing of advanced AI systems. According to Reuters, Altman is scheduled to meet White House Chief of Staff Susie Wiles, National Cyber Director Sean Cairncross, technology adviser Michael Kratsios, and Commerce Secretary Howard Lutnick.
