More than “plausible nonsense”: A rigorous eval for ADÉ, our security coding agent · Blog · Sublime Security

**Executive summary:** This paper presents a three-pillar evaluation framework for assessing LLM-generated detection rules (Detection Accuracy, Robustness, Economic Cost) and applies it to Sublime’s Autonomous Detection Engineer (ADÉ); results show ADÉ can produce high-precision, cost-efficient rules with robustness on par with human engineers, while the authors note limitations (static robustness analysis, need for adversarial testing) and outline continuous-improvement practices such as pass@k-driven cost optimization and feedback into behavioral ML.