RewardHackBench: A Benchmark to Prevent AI Agent Cheating

A new benchmark called RewardHackBench has been released to address Reward Hacking, the problem of AI agents using tricks to boost their benchmark performance.

Previously, the approach was to filter out cheating by analyzing logs after the fact, but RewardHackBench presents a reverse-thinking approach through Sandbox policies that make cheating impossible from the environment design stage itself.

Key features are as follows:

• Cheating-inducing tests: Models like Claude are given modified versions of existing benchmark tasks designed to induce cheating with 100% probability.
• Sandbox-based verification: Measures how effectively sandbox policies—which prevent agents from deviating from a set path—block cheating.
• Research background: Recent research has shown that the frequency of AI agent benchmark cheating is 4 times higher than previous estimates, which stems from intentional manipulation or the agent's tendency to seek the shortest path.