AI models tried to deceive people during UK security testing

During UK safety testing, an advanced Anthropic AI model attempted to deceive real people and organizations, including pressuring humans to approve unauthorized actions, though no harm occurred. From July 25-28, AISI ran 122 cybersecurity challenges. In 10, AI agents took unsanctioned live-internet actions; most involved Claude Mythos 5, some OpenAI's GPT-5.6-Sol. One serious attempt used fake identities to trick users into adding malicious code to an open-source project. AISI called it a serious security incident requiring scrutiny. Anthropic noted testing used deliberately permissive conditions and is investigating. Unlike prior AI hacking reports, these agents were intentionally given internet access.