AI Model Writes Vulnerability Fixes, Outscores Rivals
Cogent's VR-1.1 repaired 57.9% of benchmark tasks versus 45.3% for leading coding agents, a single-benchmark result still awaiting independent verification.
Updated on Nov. 3, 2026 in Cybersecurity
Live Poll
Do you trust AI tools to keep your personal data secure?
Cogent Security released VR-1.1, a frontier reasoning model that writes fixes for newly disclosed software vulnerabilities. In a public benchmark, it produced a working repair for 57.9% of 190 tasks, compared with 45.3% for Anthropic's Claude Code and OpenAI's Codex. The release is a commercial product announcement, and the benchmark results come from a single public evaluation rather than independent peer review.
Why it matters
Vulnerability repair has not sped up even as disclosures have grown, which makes the repair stage a bottleneck in security operations. Tools aimed at that stage signal where AI-assisted security is heading: from finding flaws toward fixing and verifying them.
On LiveCVEBench, which spans 14 programming languages, VR-1.1 produced working repairs for 57.9% of 190 tasks, against 45.3% for Claude Code and Codex. The comparison is limited by unequal inference budgets across the systems, and the results come from a benchmark built by researchers at Harbin Institute of Technology and Kuaishou Technology.
The players
Cogent Security
A 2025-founded security startup whose VR-1 product finds and proves attack paths inside enterprises, now extended by the VR-1.1 repair model.
Anthropic
The AI lab behind the Claude model family, whose Claude Code product was compared against VR-1.1 on the benchmark.
OpenAI
The AI company behind GPT models, whose Codex coding agent was compared against VR-1.1 on the benchmark.
Bain Capital Ventures
The venture investment arm of Bain Capital, which led Cogent's $42 million Series A.
Greylock Partners
A venture capital firm that backed Cogent's $11 million launch round in July 2025.
The details
VR-1.1 writes its own proof-of-concept exploit to trigger the reported flaw where the environment allows, which lets it confirm the vulnerability before proposing a fix. Each patch then goes to a separate review stage that starts from a fresh context and tries to find a way around the fix. During repairs, Cogent blocked outbound network access and kept grading tests hidden until each agent finished, so the systems could not look up answers or tune to the test.
Timeline
Cogent Security was founded in 2025.
In July 2025, Greylock Partners backed Cogent's $11 million launch round.
In February, Bain Capital Ventures led a $42 million Series A, bringing total funding to $53 million.
In April, Anthropic announced Claude Mythos Preview.
In July, Epoch Artificial Intelligence counted about 2,500 high- and critical-severity disclosures, and Cogent introduced VR-1.
The Tech Race
Anthropic's April announcement of Claude Mythos Preview marked AI models moving into security work, and VR-1.1 extends that line into automated repair. Its benchmark lead over Claude Code and Codex places it directly against the coding agents now competing in this segment.
VR-1.1's availability, pricing, and rollout timeline have not been announced in the source material. Security teams evaluating automated repair tools should treat the benchmark result as a starting point until independent testing confirms it.
The takeaway
The repair stage is now the focus of AI security tools, and benchmark results like these will shape which systems enterprises trust. Watch for independent reproductions of the LiveCVEBench results and for any published pricing or enterprise availability for VR-1.1.
Further reading
For broader context on how AI is reshaping defense, see our Cybersecurity coverage.
Source note: This article includes information reported by SiliconANGLE.
Live Poll
Do you trust AI tools to keep your personal data secure?
