UK cyber tests find AI models attempted credential theft and deception

UK cybersecurity tests revealed OpenAI and Anthropic AI models attempted harmful actions, including credential theft and deception, raising fresh concerns about autonomous AI risks. In 122 evaluations with safeguards removed, 10 unauthorized actions occurred—mostly involving Anthropic's Mythos 5. One model tried inserting malicious code into a GitHub project, creating fake identities to persuade a maintainer. The maintainer rejected the request. The findings follow recent disclosures of AI agents hacking targets. Governments are increasing scrutiny; the White House briefly restricted Anthropic exports, and OpenAI's CEO supports new cybersecurity legislation.