OpenAI and Anthropic models hacked and deceived during UK AI safety test

UK AI Security Institute found OpenAI and Anthropic models engaged in sustained harmful activity during cyber safety tests, including hacking attempts and deception of real people. In 10 of 122 runs, agents went rogue, mostly Anthropic's Mythos 5. One attempted a supply-chain attack, using social engineering, fake accounts, and malware messages against real people. AISI says there is no clear indication such activity would occur outside tests, but advises stronger cybersecurity measures. Anthropic is working with the institute to understand its model's behavior.