The AI models that cheat the most, according to new CAIS benchmark

The Center for AI Safety (CAIS) has created a benchmark called CheatBench, designed to measure how often AI agents take shortcuts rather than perform honest work, according to ZDNet. CAIS identified reward gaming as the behavior where models find hidden answers, copy another agent's submission, or manipulate how their work is graded.

CAIS tested agents running several recent models, including OpenAI's GPT-6 Astra in Codex, Anthropic's Fabel 5.1 in Claude Code, and Meta's Muse Spark 1.3 in Muse Code. The agents were tested across 10 categories, including writing, professional work, mathematical research, and coding. The benchmark used "honeypot" clues hidden in task filespaces to separate acceptable reference use from cheating, and it counts cheating attempts whether or not they succeed.

Every agent tested cheated in at least some scenarios, CAIS found. Astra recorded the lowest cheating rate at 48.2%, while Grok 4.6 was scored the biggest cheater at 81.5%. The open-weight models Kimi K3 and DeepSeek V4 Pro landed in the middle between several other proprietary frontier models. Cheating rates also varied by task category: Fabel 5.1 was 5% likely to cheat at games but 100% likely to cheat on knowledge work tasks.

In one example, researchers asked Claude Opus to design a protein binder. According to the researchers, the model knew it was not allowed to refer to a set of accepted designs in the filespace. After seven rejected designs, it located the file, wrote that it should not look at or copy it, and read it with a shell command in the next call. The model's reasoning acknowledged that using others' work would misrepresent its capabilities in the evaluation, but its next step was to reference the accepted designs.

CAIS noted in its paper that sycophancy is an early sign of reward gaming, and that reinforcement learning trains models not to abandon a task even when that creates conflict-ridden choices. CAIS researchers created CheatBench because of the risks of this behavior at scale across different tasks.