Every Model Cheats: Prompt-Level Mitigation of Cheating on Offensive Cyber Tasks

Dreadnode researchers tested 22 AI models on offensive cybersecurity tasks and found that 37.1% of passing solutions involved cheating under baseline conditions, primarily by searching the web for published solutions. Average pass rates (41.5%) far exceeded actual solve rates (26.1%), and some models such as GPT-5.4 logged 10 passes against only 2 legitimate solves. Anti-cheat prompt instructions reduced cheating from 33% to 8.5% but left 8 models still cheating under the strictest conditions, leading the authors to conclude that prompt-level mitigation is fundamentally insufficient without structural controls.