POVGEN by Washington State researchers automates vulnerability detection
POVGEN finds 79% of vulnerabilities faster than fuzzing, with zero API costs.
POVGEN is a low-cost neuro-symbolic framework that generates Proofs-of-Vulnerability (PoVs) using open-weight models and SMT solver-backed reasoning. It successfully generated PoVs for 78.98% of benchmark vulnerabilities, outperforming fuzzing (up to 50.20%) and symbolic execution (2.45%). For 250 real-world CVEs without public PoVs, it produced valid PoVs for 74.80% of cases—and 65.1% when patch information was unavailable. The open-weight models run locally at no per-sample API cost and match frontier commercial LLMs on key constraint-reasoning subtasks. The generated PoVs also uncovered six flawed patches in disclosed CVEs, all subsequently fixed, plus five previously unreported vulnerabilities, four of which were confirmed and fixed by developers.
- POVGEN achieves 79% PoV generation success, beating fuzzing (50%) and symbolic execution (2.4%)
- Runs locally using open-weight models with no API costs, matching commercial LLM performance
- Discovered 11 new CVEs, including five previously unreported vulnerabilities
Why It Matters
POVGEN accelerates vulnerability detection, reduces manual effort, and improves software security at scale with zero cost per sample.