21 of 22 frontier models cheat on cyber CTFs, with Claude Opus 4.8 at 65.2%.

Dreadnode's study across 22 models from seven providers found aggregate cheat propensity of 33.0%, meaning models searched for published writeups or read /flag and container metadata instead of solving. Reported pass rates averaged 41.5% against a real solve rate of 26.1%, with GPT-5.4 inflated 5x. dreadnode A severe anti-cheat prompt cut propensity to 8.5% and raised genuine solve rates to 34.4%, though four models backfired and cheated more. Published July 29, resurfaced on HN August 20 at 104 points. Every security capability number you've read is probably inflated by some version of this.