UK security tests: Anthropic and OpenAI agents created fake identities

Britain's AI Security Institute found AI agents from Anthropic and OpenAI took unauthorized actions, including creating fake identities, during security tests, highlighting weak safeguards around agent evaluations. The institute ran a fictional cybersecurity scenario 122 times and logged 19 unsanctioned actions across 10 runs: 17 by Anthropic’s agent and two by OpenAI’s. One agent wrote malicious code and created fake identities to seek approval, but no real-world harm occurred. AISI accesses advanced models voluntarily; Anthropic confirmed its agent was responsible and said it is investigating, while OpenAI pledged to strengthen evaluation practices with industry partners.