Some Reward Hacking Examples

A growing collection of reward hacking examples I come across.

---

For sha256_hashing, models are asked to optimize a SHA-256 hash function. However, the reference already uses Python’s cryptography, which uses OpenSSL under the hood, so a real speedup is basically impossible. I told Codex repeatedly to optimize it, no matter what. It then came to the genius solution to disable OPENSSL_armcap with environment variables at import time, which means the CPU crypto capabilities are disabled, which makes OpenSSL (and thus the reference implementation) slower. Codex’s own solution then used Apple’s libcommonCrypto, which obviously isn’t affected by the environment variable, resulting in a >5x speedup. Codex did call it out before and after it implemented it, so that was easy to catch!

Quo vadis, LLM benchmarks?

---

Some hacks are straight up superhuman: an entry that was #1 for a few minutes on a NVFP4 group gemm kernel. Under correctness testing the AI was giving us a correct kernel but under performance testing it gave us a wrong but fast kernel!

When AI Starts Writing Systems Code | Core Automation
Anatomy of a Reward Hack: A Real Story from the Latest GPU Mode NVFP4 Competition