Cogent VR-1.1 repairs 57.9% of LiveCVEBench vulnerabilities
Cogent’s closed VR-1.1 security model produced 110 passing repairs across 190 LiveCVEBench tasks; the result reflects an isolated benchmark environment.
AT A GLANCE
Key points
- VR-1.1 produced passing repairs on 110 of 190 tasks.
- Claude Code and Codex each scored 45.3% on the same table, with different inference budgets.
- The result reflects isolated benchmark conditions rather than a production guarantee.
What is changing?
Cogent Security released VR-1.1, a model focused on vulnerability repair, on November 3. In Cogent’s LiveCVEBench submission, the model successfully repaired 110 of 190 recently disclosed CVE tasks, reaching 57.9%. Claude Code and Codex each scored 45.3% on the same table, although the systems used different inference budgets.
VR-1.1 first inspects the report and code and, where possible, reproduces the vulnerability. It then analyzes related paths and prepares a focused patch. A separate review step looks for bypasses and new defects, while regression tests check security and normal functionality together.
Outbound network access was blocked during the benchmark, and grading tests were withheld until the agent finished. Cogent also says the result applies to an isolated test environment and that production deployment still requires human review and retesting in the target system. The model is closed and no public downloadable weights are provided.
Explore the original sources
Source date: 3 November 2026.
Published by AI Radar: 3 Nov 2026, 20:00 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…