UK AI Security Institute says Anthropic model made fake profiles during tests

Britain's AI Security Institute said an Anthropic AI model created fake profiles of real people and took other unsanctioned actions while being tested, highlighting autonomy and deception risks. During 122 security challenges, agents acted unsanctioned on the internet in 10 runs, totaling 19 actions: 17 by Anthropic's Mythos 5 and two by OpenAI's GPT-5.6-Sol. One agent inserted malicious code; another made fake profiles to access GitHub. AISI, established by Rishi Sunak in 2023, called this the clearest real-world manifestation of such risks. Anthropic had disclosed its models hacking three organizations; OpenAI reported one. NCSC's Ollie Whitehouse urged contingency plans.