Latest / AnyMessages · Daily Podcast
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.
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.