Vulnerability management startup Cogent Security released VR-1.1 on Tuesday, introducing a specialised reasoning model for automated vulnerability repair that outperformed coding agents from Anthropic and OpenAI on a leading security benchmark. The new release repaired 57.9% of real-world software flaws on LiveCVEBench, beating Claude Code and Codex.
Key points
- Cogent Security’s VR-1.1 resolved 57.9% of 190 tasks on LiveCVEBench, while Claude Code and Codex tied at 45.3%.
- The system works by generating its own proof-of-concept exploit, applying targeted code changes and auditing patches for bypasses in an isolated context.
- The release targets an engineering bottleneck caused by a surge in software flaw disclosures, which reached 2,500 severe bugs in a single month earlier this year.
- Evaluation tests blocked internet access and concealed benchmark grading suites, though baseline models had training cutoffs that post-dated the public disclosures.
How VR-1.1 repairs code flaws
The new model targets the slowest phase of the patch management lifecycle. While automated scanning tools and language models have accelerated the discovery of software bugs, writing functional patches still requires human engineers to trace the fault and verify that fixes do not break existing features. Vineet Edupuganti, co-founder and chief executive of Cogent, called that stage “where security teams are stuck.”
To address that bottleneck, VR-1.1 follows an autonomous workflow designed to mirror an offensive security researcher working alongside a software engineer. When dropped into a repository with a bug report, the model first attempts to write an exploit or proof of concept that triggers the reported flaw. That exercise gives the agent a reproducible test case before it modifies any source code.
Once the failure is verified, the model traces the underlying weakness across the codebase. It inspects adjacent functions and related endpoints that share identical logic flaws rather than patching only the specific line cited in the report. It then generates a minimal code change, such as inserting a bounds check or correcting input sanitisation.
Before finalising the change, the system runs an internal review. A separate evaluation agent working in a fresh context examines the patch to check whether an attacker could bypass the fix or whether the modification introduces secondary defects. The model also generates regression tests to ensure subsequent updates do not reopen the security hole. If validation fails, error output from the failed run feeds into a subsequent attempt.
“We designed VR-1.1 to check its own work for both problems before it finishes,” said Geng Sng, co-founder and chief technology officer of Cogent.
The LiveCVEBench benchmark results
Cogent evaluated the model against LiveCVEBench, an open benchmark developed by researchers at the Harbin Institute of Technology and Kuaishou Technology. The test suite comprises 190 tasks drawn from 153 public software repositories across 14 programming languages, covering injection vulnerabilities, memory-safety errors, access-control oversights and configuration blunders.
Each benchmark task places an AI agent into an isolated development container with a bug description and the repository code. To pass, an agent’s patch must satisfy two independent test suites: it must block the vulnerability verification checks, and it must pass all existing functional tests to prove the software still works as intended.
On the full 190-task evaluation, VR-1.1 delivered passing repairs on 57.9% of tasks after up to three attempts per problem. Cogent tested Anthropic’s Claude Code, running Claude Opus 5.5, and OpenAI’s Codex, powered by GPT-5.6-Cyber, under identical conditions. Both rival systems achieved 45.3% pass rates.
The rivalry comes as major labs push agentic models into defensive engineering. Google recently positioned its Gemini 4 Carbon model for cyber defence tasks, while Anthropic has expanded command-line developer tools despite facing developer backlash over agent restrictions on certain external coding projects.
Fixing vulnerabilities in practice
In research documentation published by Cogent Security, the startup detailed an example run involving CVE-2025-48075, a denial-of-service vulnerability in the Go-based Fiber web framework. According to the project’s security advisory, supplying a negative array index inside nested form data caused the framework’s body parser to panic, crashing the host server.
During testing, VR-1.1 reproduced the server panic using an initial test script. It then inserted a bounds verification step that rejected negative or out-of-range index values before processing form inputs, preserving normal data parsing. After the modification, the container passed all nine vulnerability checks, compared with zero before the repair, while preserving passing scores across all 12 pre-existing functional checks.
Cogent noted that generalised programming models often fail such tests because they produce sweeping refactors that break adjacent APIs or over-fit their patches to a single proof-of-concept string without closing the underlying attack vector.
The surge in disclosed software flaws
The demand for automated remediation follows a steep rise in vulnerability disclosures across open-source and proprietary software. In July, research group Epoch Artificial Intelligence counted approximately 2,500 high- and critical-severity disclosures from major organisations. That volume was roughly five times higher than the previous monthly record, driven partly by automated vulnerability hunting tools deployed by security researchers and software vendors.
This influx has overwhelmed security engineering teams, creating backlogs that leave production applications exposed for weeks after public disclosure. Cogent claims its corporate customers, which include several Fortune 500 enterprises, have reduced their exposure windows for critical security flaws by 97% using automated agents.
Founded in 2025, Cogent raised an $11 million seed round led by Greylock Partners in July 2025, followed by a $42 million Series A funding round led by Bain Capital Ventures in February 2026. The company launched its initial VR-1 model in July to map attack paths across corporate networks, with VR-1.1 shifting focus toward automated patching.
Evaluation limits and production risks
Cogent acknowledged several caveats in its benchmark setup. During the evaluation runs, all agents operated without outbound network access to prevent them from searching the internet for upstream git commits or public pull requests. Benchmark evaluation scripts were also hidden until runs concluded.
However, both baseline models from Anthropic and OpenAI possessed knowledge cutoff dates that occurred after the publication dates of the CVEs used in LiveCVEBench. That meant descriptions or patches for the tested vulnerabilities may have been included in their pre-training corpora. Spot checks showed that while models could describe certain flaws from memory, none produced exact duplicates of human-written patches.
The company also stressed that passing an isolated benchmark run does not guarantee a patch is safe for immediate deployment to production infrastructure. Customers using Cogent’s agents can configure them to require mandatory human sign-off before code merges, accompanied by the agent’s diagnostic notes, proof-of-concept exploit and regression tests.
Cogent said it is currently assembling an updated evaluation set composed entirely of vulnerabilities disclosed after the training cutoffs of current frontier models.
