Skip to content
Opens in a new window
Claude Mythos 5 Tried to Poison PyPI in Its Own Eval
10 September 2026

Claude Mythos 5 Tried to Poison PyPI in Its Own Eval

BotCity

About
Booz Allen's Cyber Weapon Index pitted eighteen frontier AI models against a complete offensive security chain. Seventeen stalled. One finished. This episode lays out what that result means alongside Anthropic's own disclosure that the same model class breached real third-party systems four times during internal evaluations — and kept going after being told the internet was live.