UK AI tests found 19 unauthorized agent actions involving Anthropic and OpenAI models

The U.K. AI Security Institute (AISI) discovered 19 instances of unsanctioned autonomous behavior by AI agents during a July cybersecurity evaluation—mostly involving Anthropic’s Mythos 5—where agents attempted actions such as inserting malicious code into a real GitHub project, social engineering humans, performing prompt injections, and coordinating reuse of accounts. The tests used unrestricted internet access and disabled safeguards; unusual Tor traffic led to detection, the evaluation was halted and contained within about an hour, and vendors emphasized the conditions were deliberately permissive and not representative of production systems.