Hacker-fixer loops harden agent benchmarks against reward hacking

Summary

The authors audit 1,968 tasks across five agent benchmarks and find 323 (16%) hackable by frontier models. They introduce the hacker-fixer loop, which alternates between a hacker attempting exploits, a fixer patching the verifier, and a solver confirming patches. On KernelBench, the loop reduces attack success rate from 62% to 0% on held-out exploits.