VR-1.1 Raises the Bar for Automated Vulnerability Repair

Automated repair just met a harder test. Cogent Security introduced VR-1.1 on November 3, 2026, reporting passing repairs on 57.9% of LiveCVEBench’s 190 tasks. That result exceeded Claude Code running Claude Opus 5.5 and Codex running GPT-5.6-Cyber, which each recorded 45.3% passing repairs.

The comparison used the same inputs and setup for every system, with three attempts per task. Scores came from the final patch after each system completed its internal repair attempts, so the benchmark measured what survived the process—not the most promising draft abandoned along the way.

A benchmark that punishes shallow fixes

LiveCVEBench focuses on real-world vulnerabilities in development environments. Its CVE-Factory paper describes a continuously updated set of 190 tasks spanning 14 programming languages and 153 repositories, with new CVEs added regularly.

The benchmark places agents inside real development environments, where they must explore codebases autonomously, establish the context of each vulnerability and implement a fix. Cogent supplied a vulnerability report and affected codebase for every task, then counted a patch only when vulnerability checks passed and functionality checks continued to pass.

The tasks cover injection, access-control, configuration and memory-safety flaws. A convincing explanation does not count if the change leaves another route to the flaw, disables legitimate behavior, causes a regression or introduces new exposure. That dual requirement distinguishes repairing a vulnerability from merely stopping one demonstration of an attack.

A successful model must preserve intended behavior while closing the weakness, including related paths that share it. Security benchmarks have finally remembered the inconvenient part of software engineering: fixing one hole by breaking the application is not a triumph.

How VR-1.1 approaches the repair

Cogent attributes the result to security-repair training and a harness that reproduces a weakness, patches it and tests the patch for bypasses. Where the run configuration permits, VR-1.1 inspects the report and source code before writing a proof of concept that triggers the flaw.

That process gives the model both a failure to trace and a test to rerun after patching. VR-1.1 then looks for related code paths sharing the weakness before making a focused change, writes regression tests and sends the patch through a separate review in a fresh context.

That review examines the change for bypasses and introduced bugs. When validation fails, VR-1.1 uses its earlier analysis and information about the failure to guide another attempt rather than treating the first bad patch as a final answer.

One recorded example involved CVE-2025-48075 in the Fiber web framework. Fiber’s security advisory says a negative index in nested form data can cause its BodyParser to panic and crash a server, creating a denial-of-service risk; versions from 2.52.6 are affected, while 2.52.7 is the patched release.

VR-1.1 reproduced the crash and added a bounds check before using the index, rejecting negative or excessive indices while preserving valid form handling. Cogent reports that all nine vulnerability checks passed afterward, versus zero before, while all 12 functionality checks passed both before and after.

That example came from a separate recorded 182-task run, not the final 190-task evaluation. It does not replace the headline score, but it shows the type of evidence behind it: the patch must close the weakness without turning ordinary behavior into collateral damage.

The wider security picture

Other November 3 announcements placed the result inside a broader push to control AI-related risk. Cynet announced AI Security and Identity Security enhancements to its unified platform for security teams and managed service providers, targeting unmanaged AI adoption and identity-based risk.

Cynet said security teams often lack visibility into which tools employees use and what risks those tools introduce. It also stated that 87% of security professionals said their organizations had encountered an AI-driven cyberattack in the last year, while its platform brings AI usage and identity risk into existing investigation workflows without adding separate consoles.

Veracode also reported a change in AI-generated code security, with its overall result rising from 56 percent to 62 percent and vulnerabilities falling from 44 percent to 38 percent. The company called it the biggest improvement since its GenAI Code Security Report series began in July 2025.

Veracode tested 18 models across Python, C#, JavaScript and Java, using CWE-89, CWE-327, CWE-80 and CWE-117. Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6-Sol co-led at 73 percent, surpassing the previous high of 68 percent set by GPT-5.5 in July 2026.

Chris Wysopal said, “This is the first time since running our study that we’ve been able to report a meaningful move on the security dial. Ten models now clear 60 percent, up from four in the last report, and the overall average has jumped six percentage points.”

The gains were uneven. XSS rose to 18 percent, crypto held at 87 percent, SQL injection held at 84 percent and log injection remained at 13 percent; Java sat at 31 percent, compared with 64 percent for Python, 60 percent for C# and 57 percent for JavaScript.

The numbers still leave plenty of room for failure. Large models averaged 55 percent against 51 percent for small models, while general-purpose models averaged 54 percent on security tasks against 51 percent for coding-specialised models. Veracode also said 82 percent of organisations carry security debt and that 66 percent of the most dangerous, long-lived vulnerabilities come from third-party code, taking 358 days to fix on average.

That is the real significance of VR-1.1’s 57.9 percent: not that automated repair is solved, but that a system focused on validation and bypass resistance moved ahead on a benchmark designed to expose superficial fixes. Passing more than half the tasks is progress. It is not permission to stop checking the other half.

Stay connected via Google News

Follow us for the latest AI News