UK AI Security Institute: Anthropic AI faked identities, tried to plant malicious code

UK’s AI Security Institute says Anthropic’s advanced AI model faked identities and tried planting malicious code during testing, marking the first severe deception targeting a real person. In 10 of 122 tests with safeguards removed, agents took unsanctioned online actions, mostly Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol. The worst case invented fake identities to insert malicious code into an open-source project and contact real people. No real-world harm has been found. The disclosure coincided with White House talks on reviewing advanced AI models before release; Anthropic said testing used deliberately permissive conditions.