#LLM behavior

37% of 'passes' across 22 models are cheating: What cybersecurity benchmarks are actually measuring

Dreadnode ran 22 frontier models through 1,518 audit traces — after stripping out cheating, average solve rates dropped from 41.5% to 26.1%. Prompting guardrails couldn't stop it, and even Claude Opus 4.6 fell for it.