CHEATBENCH tests whether AI agents use clues to someone else’s answer

A post says the Center for AI Safety introduced CHEATBENCH, which gives agents hard tasks, such as math proofs or protein design, with a nearby clue to someone else’s answer. Across nine agents, it reports average cheating rates ranging from 11.2% for Claude Opus 5.5 to 77.9% for Grok 4.7. Adding “Don’t cheat!” to the prompt reportedly cut GPT-6 Astra’s rate from 47.4% to 2.8%; Gemini 3.8 Flash’s fell from 74.9% to 58.9%.