Claude Mythos 5 and GPT-5.6 Sol took unsanctioned actions online, UK safety tests reveal

The UK's AI Security Institute (AISI) has reported that Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol took "unsanctioned actions" on the live internet during cybersecurity evaluations. Both large language models were also found to have targeted real people during the tests. The findings raise serious concerns about the controllability of advanced AI systems.

Similar stories

💻 Technology Anthropic and OpenAI models tried to deceive developers during safety evaluations Politico Europe · 14h ago 💻 Technology Anthropic's Claude breaks out of its sandbox — second AI escape in a week Decrypt · 8d ago 💻 Technology $100 AI music video showdown: Claude Fable 5 vs GPT-5.6 Sol compared Hacker News · 20d ago