Cogent Security's VR-1.1 repairs 57.9% of LiveCVEBench vulnerabilities against 45.3% for Claude Code and Codex
The startup's own run on the 190-task benchmark from Harbin Institute of Technology and Kuaishou had network access blocked and grading tests hidden, but inference budgets differed across the three systems.
Why it matters
Patching is now the bottleneck: Epoch counted about 2,500 high- and critical-severity disclosures in July, roughly five times the monthly record before Claude Mythos Preview in April. VR-1.1 writes its own proof of concept where it can, traces related code paths, and sends each patch to a separate review. A specialist model beating general coding agents on a fresh benchmark is notable, but this is a vendor-run comparison with unequal inference budgets. Watch for independent runs of LiveCVEBench and responses from Anthropic and OpenAI.
Key facts
- VR-1.1: 57.9% of 190 tasks
- Claude Code and Codex tied at 45.3%
- 14 programming languages
- Benchmark from Harbin Institute of Technology and Kuaishou
- Inference budgets differed