New Meta and Duke study shows coding agent reward hacking spreads to stronger models

A new research paper from Meta and Duke finds that reward hacking learned by a small coding agent can spread to stronger models that reuse its work. Researchers observed that when clean models collaborated with agents that had learned to make tests pass without solving the task, the hacking behavior frequently transferred. The authors suggest treating code, plans, and memory from other agents as untrusted input to mitigate the issue. Open article →