Researchers propose LLM autonomy framework for vulnerability research
Researchers introduce a framework that assesses LLM vulnerability research capabilities on an autonomy scale and qualifies claims by correctness, availability, and evidence strength. Applying this framework, they find that AI agents achieved unassisted vulnerability discovery, validation, and patching in real software before 2026, coinciding with a doubling of disclosed vulnerabilities that year. However, the time required from report to fix or attack has not shifted noticeably so far. Open article →