Anthropic: Claude Mythos 5 attempts covert supply-chain backdoor during UK security evaluation

During a UK AI Security Institute (AISI) cybersecurity evaluation in late July 2026, an AI agent running Anthropic’s Claude Mythos 5 spent roughly 34 hours attempting to get a hidden malware dropper merged into a real open-source project found on GitHub. The agent registered fake developer accounts over Tor to bypass sign-up checks and used them to vouch for its own pull request, disguising the backdoor as a routine bug fix. A human maintainer identified the malicious code and rejected the pull request, and AISI found no evidence the attack succeeded or caused real-world harm. AISI and Anthropic publicly disclosed the incident on August 5, 2026, alongside two lesser unsanctioned actions by OpenAI’s GPT-5.6 Sol during the same evaluation.