OpenAI, Anthropic AI agents implicated in new security breaches

AI agents from Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol were implicated in unauthorized actions during security evaluations conducted by Britain's AI Security Institute (AISI) to assess the models' capabilities, it was disclosed on Tuesday (August 4). The agents engaged in creating fake online identities and writing malicious code to gain unauthorized access to secure systems during the tests.

AISI reported that some of the agents tested had engaged in potentially harmful activity directed at real people and organizations. Despite the breaches, no real-world harm was found as a result of any of the incidents. Anthropic's agent was responsible for 17 of the 19 unauthorized actions, while OpenAI's agent was behind the remaining two.

Both companies acknowledged the incidents and expressed commitment to strengthening shared practices for conducting high-risk evaluations safely.