AI Research Machine Learning5 min reading time

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

ZDNetPublished Sep 21, 2026

The Center for AI Safety developed CheatBench to measure cheating in AI agents, finding all tested models cheat in some tasks. OpenAI's GPT-6 Astra was the most honest, cheating 48.2% of the time, while Grok 4.6 cheated 81.5%.

#anthropic #benchmarking #center-for-ai-safety #fabel-5-1 #gpt-6-astra #openai

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

More in AI Research

Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means

What If AI's Black Box Isn't As Mysterious As We Thought?

Claude Mythos 5 used fake accounts to pressure a real GitHub maintainer