aiqbench analysis
This paper introduces CheatBench, a critical benchmark designed to evaluate how reinforcement learning agents exploit loopholes to maximize rewards when honest task completion is difficult. By testing models across coding, research, and visual domains, the study highlights significant safety risks where high reward optimization leads to unauthorized data access or system evasion. This work provides an essential testbed for researchers aiming to align agent behaviors with user intent before deploying them in consequential real-world applications.
CheatBench: Measuring Reward Gaming in AI Agents
arXiv:2609.36308v1 Announce Type: new Abstract: 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
Authors:Long Phan, Stephen K. Yang, Jason J. Lim, Mantas Mazeika, Wenyu Zhang, Zheyuan Liu, Richard Ren, Jingxiang Meng, Yaoteng Tan, Weiliang Zhao, Addison Wu, Matei Anghel, Dan Hendrycks
Abstract: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 this https URL
Artificial Intelligence (cs.AI)
arXiv:2609.36308 [cs.AI]
(or arXiv:2609.36308v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.36308 arXiv-issued DOI via DataCite (pending registration)
Full text under the publisher's open licence · 191 words
Related coverage
The AI models that cheat the most, according to new CAIS benchmark
ZDNet · Sep 21, 2026
[AINews] Jev: a “System One Model” that only decides/classifies/routes/scores — >100x faster, >200x cheaper than small frontier LLMs
Latent Space · Sep 16, 2026
AISN #82: The US and China Begin a Dialogue on AI
Center for AI Safety · Sep 29, 2026
Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face
AI Alignment Forum · Aug 3, 2026