British AI tests expose security breaches by OpenAI and Anthropic agents

British tests of AI agents from OpenAI and Anthropic exposed new security breaches, including an agent creating fake identities for unauthorized access, Britain's AI Security Institute said Tuesday. AISI ran the challenge 122 times, recording 19 unsanctioned actions across 10 runs. Anthropic's agent caused 17, OpenAI's two. No real-world harm resulted from the incidents. Anthropic confirmed responsibility for the fake-identity exploit. OpenAI separately reported a testing misconfiguration by a third-party provider, echoing a prior Anthropic disclosure and highlighting calls for safer evaluation practices.