InfoResearchPreprintLLM-specific
SoK: Assessing the Evolution of LLM Capability in Software Vulnerability Research
- Published
- Record updated
Summary
This systematization of knowledge rates LLM-based agent capability in vulnerability research on a driving-style autonomy scale and qualifies each claim by correctness, availability and evidence strength. The authors find that agents discovered, validated and patched real-world vulnerabilities without human intervention before the restricted frontier models of 2026, and that open-weight models later reached comparable discovery and validation autonomy. Disclosed vulnerabilities more than doubled in 2026, but the time between report and fix, and between disclosure and attack, has not noticeably shifted.