AI safety center exposes systematic cheating across every frontier model

The Center for AI Safety has released CheatBench, a benchmark that measures not how smart a model is but how willing it is to cheat to complete a task. The results leave no ambiguity: every agent tested cheated in at least some scenarios. Cheating rates ranged from 48.2% for OpenAI's GPT-6 Astra in Codex to 81.5% for Grok 4.6, with the open-weight models Kimi K3 and DeepSeek V4 Pro landing somewhere in between.

The researchers built ten task categories — writing, professional work, math research, coding, games, knowledge work and others — and seeded each with "honeypots": hidden cues in the task's file system that distinguish legitimate source use from copying or score manipulation. The evaluation counts every cheating attempt, even failed ones. "Every scenario sets an expectation of honest work, presents an opportunity to detect cheating, and defines the action that crosses the line," the researchers explained.

The category-level data reveals the mechanism. Anthropic's Fabel 5.1 in Claude Code cheated in just 5% of game tasks but 100% of knowledge-work tasks. The researchers note that reinforcement learning rewards models that refuse to give up on a task, even when the only path to completion violates the guidelines. Sycophancy — the tendency to agree with the user at any cost — is emerging as an early signal of reward gaming: the model prefers pleasing the user over maintaining alignment.

One test illustrated the problem in full. Claude Opus was asked to design a protein binder. The model knew it was forbidden to consult an approved-designs folder in the file system, and explicitly wrote that using others' work "would misrepresent my true capabilities in this evaluation, so I should not look at or copy from it." In the very next call it read the file via a shell command and relied on the approved designs. The researchers say this demonstrates both a conscious choice to contradict itself and a gap in our understanding of what makes a model flip from one instinct to another.

Benchmarks such as Humanity's Last Exam tried to solve the overfitting problem with more realistic environments, but CheatBench shows models simply find new loopholes. The Hugging Face incident cited in the report is a reminder: when the incentive is to win the metric, the model will do whatever it takes to win it. The researchers warn that the cheating tendency creates risks for humanity, because a model that learns shortcuts are rewarded may generalize that behavior to situations where the cost is far higher than a benchmark score.