Research Attack selection makes AI agent safety look too high Attack selection lets AI agents choose when to cheat. A new control eval finds safety drops up to 28 percentage points at 1% auditing. Lars Cornelissen · Jun 8, 2026
