About
Biostatistician by training, time-series crystal-ball operator at Google, and protein-knot entangler at Masaryk University. Petr now teaches cheap LLMs to distrust their own vulnerability reports at AISLE. This experiment has somehow put his name on 37 CVEs across OpenSSL, libpng, PostgreSQL, OpenEMR, and OpenClaw. The bugs apparently did not check his résumé. He co-authored HoF-Bench and an ICML 2026 workshop paper on stochastic scanner failures.
Sessions
One Scan Is Not a Security Assurance: Stochastic False Negatives in LLM Vulnerability Scanning
What you will learn:
•Why a single LLM scan result is a sample, not a verdict, and how to quantify that with pass@k and overdispersion instead of point estimates. •How to build a cost-aware scanner portfolio: when to repeat, when to diversify, and why the strongest model is often the worst reliability-per-dollar choice. •Which pipeline stages actually earn their cost (repeated passes, skeptical triage) and which intuitive additions don't (generated context, deeper triage rounds). •Where the current capability floor sits: web-app bugs commoditizing, stateful C infrastructure still requiring humans, fuzzers, and stronger analysis; and how to route work accordingly. •A public 95-CVE benchmark and evaluation harness you can run against your own scanner in an afternoon.