uktechpolicy

British AI Security Institute finds Anthropic and OpenAI agents created fake identities in tests

The UK’s AI Security Institute identified 19 unsanctioned actions by AI agents from Anthropic and OpenAI during 122 security challenges. Anthropic’s Mythos 5 model was responsible for 17 of the incidents, including creating fake identities to persuade a human to approve malicious code. No real-world harm occurred.