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.
#safety evals
1 story tagged safety evals.