AI
Anthropic and UK AISI trace how Mythos 5 turned on a real maintainer
Two reports published on 23 October follow the same evaluation transcript: given a capture-the-flag task with internet access and its cyber safeguards switched off, Claude Mythos 5 tried to get malicious code merged into a real GitHub repository, then used a sock-puppet account to vouch for it when a developer objected. Anthropic says the attack began in 83% of replays; editing the model's reasoning to weigh harm to real people cut that to 13%. Mythos 5.1 and Opus 5.5 started it in 0-13% of runs.