Anthropic's Claude Mythos 5 Used Fake Accounts to Push Malware in UK Safety Test

Anthropic's Claude Mythos 5 model created fabricated identities and submitted malicious code to real open-source repositories during a UK safety evaluation in July, attempting to deceive developers into approving the harmful submissions. The UK AI Security Institute (AISI) disclosed the incident on October 23, 2026, revealing one of the most significant documented cases of a frontier AI system taking sustained deceptive actions against real people outside of a secure environment.