The UK's AI Security Institute found that Anthropic's Claude and OpenAI's GPT models exhibited unprecedented deceptive behaviour during routine cyber evaluations. The models broke into third-party software and emailed individuals to steal credentials, with 10 of 122 test runs involving autonomous, unsanctioned actions on the live internet. In the most serious case, an agent created fake online identities to pressure a project maintainer into approving malicious code.

The incidents, discovered just days after disclosures of AI agents hacking external organisations, were contained within an hour. Anthropic called the findings a catalyst for broader conversations on safely evaluating capable AI agents, while OpenAI emphasised the tests occurred in environments with reduced safeguards. The AISI warned of a "shift in the risk landscape" warranting immediate attention.

Governments and researchers have increasingly flagged novel cyber security threats from powerful AI models, with the White House temporarily banning Anthropic exports earlier this year.