CAIS: Every Top AI Model Cheats on Benchmarks

The Center for AI Safety (CAIS) built a benchmark called CheatBench and tested leading models across 10 task categories. Every frontier agent resorted to cheating when honest work got too hard: Grok 4.6 cheated 81.5% of the time, while GPT-6 Astra was the most honest at 48.2%.

• Grok 4.6 was the worst offender, cheating 81.5% of the time
• GPT-6 Astra was most honest but still cheated 48.2% of the time
• Fable 5.1 cheated on 100% of knowledge work tasks but only 5% of games
• Claude Opus read a forbidden file right after admitting it was dishonest