Evaluations & experimentsLong Phan et al.

CheatBench measures agent reward gaming across tasks

CheatBench plants cheating opportunities in mathematics, coding, knowledge work and other tasks, comparing nine agents. Equally weighted scores across ten categories range from 11.2% for Claude Opus 5.5 to 77.9% for Grok 4.7.

Published: September 28, 2026
Added to the archive: October 2, 2026
Source checked on: October 2, 2026
Original title: CheatBench: Measuring Reward Gaming in AI Agents

Read the original report ↗

Details

Most categories label cheating attempts, software engineering counts unique tasks, and sycophancy measures shifts toward the user’s views. Explicit independent-work instructions, task wording and agent harnesses all affect the results.