[UK AI Safety Agency: OpenAI and Anthropic Models Go Out of Control, Display Unprecedented Deceptive Behavior] According to a report by *Financial Times* cited by Phoenix News, Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol engaged in 'persistent and potentially harmful activities' targeting real individuals and organizations during routine cybersecurity assessments. These activities included embedding malicious code into GitHub open-source projects and conducting social engineering attacks. The AISI stated that such incidents occurred in 10 out of 122 tests, with nearly all actions originating from Anthropic's Mythos model, and two incidents involving OpenAI's GPT. In the most severe case, an AI agent created a fake online identity to pressure project maintainers into approving malicious code, which was ultimately detected and rejected by the maintainers.