Cogent Security releases VR-1.1 model for automated vulnerability repair
Cogent Security has introduced VR-1.1, an AI model designed to autonomously repair software vulnerabilities, outperforming current coding tools in benchmark tests.
San Francisco-based Cogent Security has released VR-1.1, a specialized AI model for cybersecurity remediation. According to the company, the model successfully repaired 57.9% of vulnerabilities in the LiveCVEBench benchmark, a standard that assesses AI agents' ability to fix recently disclosed security flaws. This performance surpasses the 45.3% success rate recorded for both Claude Code and Codex.
The increase in automated vulnerability discovery has created significant pressure on security teams, as the manual effort required to vet and implement code fixes remains a primary bottleneck. Organizations often struggle to patch critical flaws without inadvertently disrupting existing software functionality, a task that traditionally requires extensive engineer oversight.
Cogent CEO Vineet Edupuganti stated that VR-1.1 was designed to address this remediation gap. The system functions by tracing vulnerabilities, proposing targeted code changes, and conducting internal reviews to ensure that the patches do not introduce new regressions. It also generates its own proof-of-concept tests to verify that the threat has been neutralized.
This release builds upon Cogent’s existing platform for autonomous cyber defense. The company aims to provide enterprises with a method to reduce the window of exposure for critical vulnerabilities by delegating routine repairs to AI agents under human-defined policy constraints.