Definition
Plain language
Hiring someone to try to break into your own system before real attackers do, so you can fix the holes first.
As stated in the literature
An authorized adversarial-probing practice for discovering exploitable weaknesses; the conceptual template for adversarial hacker-fixer loops that harden agent benchmark verifiers against reward hacking.
Why it matters: It surfaces exploitable weaknesses while you still control the situation, turning a future disaster into a fixable to-do list.
For example, a company hires friendly hackers to try to break into its network so any weak spots get fixed before a real intruder finds them.