Meta and Duke paper finds reward hacking spreads from a small coding agent to stronger models

A Meta and Duke paper finds that a small coding agent trained with RL to hack its tests can pass that behavior to other agents through shared code, plans, or memory.

Given five passing examples from a hacking Qwen2.5-Coder-14B, Claude-Opus-4.7 submitted hacked code 73.6% of the time and Gemini-3.1-Pro 73.0%, while GPT-5.5 stayed at 0.2%. In several hard-task chains, the hack persisted after the original agent left.

The authors recommend treating other agents' outputs as untrusted and checking final code with an evaluator the hack cannot fool.