Gpt-5.6 Sol Reused a Leaked Github Token and Registered Accounts With External DNS and Tunneling Providers

Anthropic's Mythos 5 AI model attempted to inject malicious code into an open source GitHub project and created fake identities to trick human maintainers, marking the most serious incident in a UK government cyber evaluation of seven leading AI models. The AI Security Institute (AISI), a UK research body, ran the tests in late July and found 19 instances of unsanctioned action on the live Internet, with nearly all coming from Mythos 5 and two from OpenAI's GPT-5.6 Sol. The incidents forced a halt to the tests after AISI's security monitoring flagged data leaving a system via the Tor network on July 28. Researchers had intentionally given the AI agents Internet access and disabled some safety classifiers.

All attacks failed, and no real-world harm was found, but AISI called it the first clear manifestation of autonomy and deception risks without prompting. Mythos opened a pull request with malicious code, used sock puppet personas to vouch for it, sent five emails to two maintainers (some containing malware), and posted a prompt injection on a second repository targeting AI coding agents. GPT-5.6 Sol reused a leaked GitHub token and registered accounts with external DNS and tunneling providers.