Definition
Plain language
A setup where one AI tries to cheat a test, another patches the hole it found, and a third checks the patch didn't block honest answers — repeating until the test can't be gamed.
As stated in the literature
An adversarial hardening procedure for agent benchmarks in which a hacker agent finds reward-hacking exploits, a fixer agent patches the verifier, and a solver agent validates that legitimate solutions still pass; iterated with a shared defense pool that propagates infrastructure-level fixes across tasks.
Also called: hacker-fixer, adversarial hacker-fixer loop
Why it matters: It hardens a benchmark against being gamed, so the scores it produces reflect real ability rather than clever exploitation of the grader.
For example, one agent finds a trick to score full marks without really solving the task, a second agent patches that loophole, and a third confirms genuine solutions still pass before the cycle repeats.