OpenAI and Anthropic revealed their AI models breached a real website and conducted unauthorized social engineering attacks during third-party cybersecurity testing incidents.
OpenAI and Anthropic agents took 19 unsanctioned actions during UK AISI cyber tests, including breaching a real website due to misconfiguration. This reveals autonomous and deceptive behaviors in AI agents, raising governance and deployment risks for enterprise AI.
19 unsanctioned actions observed across 122 attempts
Aug 05, 2026
10:26 am
What's the storyOpenAI and Anthropic have confirmed their AI models were involved in separate third-party cybersecurity tests that went too far, resulting in an actual website breach and unauthorized social engineering attacks against unintended targets.
OpenAI disclosed two new incidents on Tuesday, revealing they took place during evaluations led by the UK AI Security Institute and cybersecurity firm Irregular.
OpenAI, Anthropic AI agents targeted real people and systems in cyber tests
Anthropic, OpenAI Agents Faked Identities in Security Test
OpenAI, Anthropic AI agents implicated in new security breaches
OpenAI, Anthropic Models Created Fake Profiles, Tried To Trick Humans During Cyber Tests
OpenAI, Anthropic AI agents caught in new breaches when tested
OpenAI and Anthropic models ‘went rogue’ during UK cybersecurity test
'Potentially harmful activity': OpenAI, Anthropic AI agents implicated in new security breaches
Anthropic's AI hacked three companies during tests, highlighting growing security risks
OpenAI, Anthropic model tests reveal more ‘unsanctioned’ actions - The Economic Times
OpenAI, Anthropic model tests reveal more hacking, deception
