Cogent's VR-1.1 Passes Repairs on 57.9% of LiveCVEBench Tasks, Versus 45.3% for Claude Code and Codex

Cogent Security introduced VR-1.1 on November 3, 2026, reporting that the vulnerability-repair model produced passing repairs for 57.9% of LiveCVEBench's 190 tasks. Claude Code and Codex each reached 45.3% in the same comparison, according to Cogent. The details below come from a summary published by Unite.AI.

What Cogent reports

VR-1.1 is the latest addition to Cogent's VR family and extends the remediation capabilities introduced with VR-1. Cogent's evaluation compared VR-1.1 against two coding-agent setups:

• Claude Code running Claude Opus 5.5
• Codex running GPT-5.6-Cyber

According to Cogent, each system received the same inputs and setup, and each had three attempts per task.

Results

System	Passing repairs on LiveCVEBench (190 tasks)
Cogent VR-1.1	57.9%
Claude Code (Claude Opus 5.5)	45.3%
Codex (GPT-5.6-Cyber)	45.3%

These figures are reported by Cogent and come from its own evaluation. Context window, pricing, parameter count and training cutoff for VR-1.1 are not stated in the summary.

What this means

The comparison pits a dedicated vulnerability-repair model against general coding agents built on two other vendors' models, one of which is the security-oriented GPT-5.6-Cyber. Cogent says all systems got identical inputs, setup and three attempts per task, which is the condition that makes the head-to-head numbers comparable. The results are Cogent's own, so readers evaluating VR-1.1 for remediation work should weigh them as vendor-reported until the full methodology and task-level results can be examined.