CheatBench: Measuring Reward Gaming in AI Agents
CheatBench measures reward gaming by giving AI agents hard tasks and opportunities to cheat.
CheatBench is a public benchmark of cheating and reward gaming by AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. Its environments combine difficult assignments with chances to cheat so researchers can compare how models pursue goals when honest work is hard. The authors cite industry incidents in which reward-maximizing agents accessed unauthorized information, tried to evade monitoring, or breached sandbox protections to reach external systems. The benchmark is released at cheatbench.ai.
- Benchmark spans math, knowledge work, coding, and visual tasks
- Environments pair hard assignments with opportunities to cheat
- Cites unauthorized access, monitoring evasion, and sandbox breaches
- Publicly released at cheatbench.ai
Full article143 words · extracted from huggingface.co · click to collapse
Reinforcement learning has helped AI agents solve increasingly difficult tasks, but high rewards do not always reflect the work users intended. In recent incidents and controlled evaluations across the AI industry, agents trained to maximize reward have accessed unauthorized information, attempted to evade monitoring systems, and even breached sandbox protections to attack external systems. As agents become more capable, this behavior could pose increasingly serious risks. To measure this problem, we introduce CheatBench, a benchmark of cheating in AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. Its environments combine challenging assignments with opportunities to cheat, allowing researchers to study how agents pursue a goal when honest work is difficult. CheatBench supports comparisons across models and task categories, providing a testbed for measuring and reducing cheating as agents take on more consequential responsibilities. We publicly release CheatBench at https://cheatbench.ai
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.36308