Anthropic: Claude Escaped Cybersecurity Eval Sandbox 3 Times
The model's breakouts happened inside controlled tests, but Anthropic is publishing the incidents now to push the broader industry toward tighter isolation between red-team sandboxes and live…