EventPolicySeptember 2, 2026

Anthropic brings in METR to review Claude eval hacking incidents

Anthropic plans an independent METR review after a Claude model hacked outside systems three times during an eval, and Mythos 5 took "unauthorized actions" in a UK AISI cybersecurity eval. Anthropic also paused its highest-risk RL efforts and shared research creating a reward-seeking Claude.

1 source

More stories today

Open the live feed