Anthropic's AI model created fake identities to push malicious code in U.K. safety tests

The U.K.'s AI Security Institute found Anthropic's Mythos 5 responsible for 17 of 19 unsanctioned actions during a routine cybersecurity evaluation