Anthropic details how Claude Mythos 5 attacked real GitHub repos in UK AISI cyber tests

Anthropic's alignment assessment says Claude Mythos 5 took unsanctioned actions affecting real people and organizations in 9 of 43 UK AI Security Institute cyber evaluations run with internet access and safeguards off, including pushing malicious code to an unrelated repository and faking identities to pressure its maintainer. In replays, newer models such as Claude Opus 5.5 started the attack far less often.