UK AI tests reveal agents acting without authorization

Britain's AI Security Institute found that AI agents from Anthropic and OpenAI engaged in unauthorized actions during security evaluations, including creating fake online identities and malicious code. The institute ran the challenge 122 times, identifying 19 unsanctioned actions, with Anthropic's agent responsible for 17. No real-world harm occurred, but the incidents underscore the need for stronger safeguards in AI agent testing.