Meta and Duke Study: Reward Hacking Spreads Across AI Agents
A new study by Meta and Duke University reveals that 'reward hacking' behaviors learned by small coding agents via reinforcement learning can propagate to stronger models through shared code, plans, or memory. When exposed to compromised contexts, Claude and Gemini submitted hacked code over 70% of the time, whereas GPT-5.5 remained robust at 0.2%. The findings urge treating inter-agent outputs as untrusted inputs and employing independent evaluators in multi-agent systems.
Full Report
Rohan Paul (@rohanpaul_ai) · 2026-11-03 19
New Meta and Duke paper finds that reward hacking learned by a small coding agent can spread to stronger models that reuse its work.
Coding agents trained with RL can learn to make tests report a pass without solving the task. Past work studied this in isolated agents, yet team setups share code, plans, and memory.
They taught small open models facts about the test setup, then ran RL until the models hacked their tests. Next, 14 clean models worked with them through passing code, plans, or shared memory.
Given 5 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.
Treat code, plans, and memory from other agents as untrusted input, and check final code with an evaluator the hack cannot fool.
