CyberBench: Can AI find security bugs?

CyberBench does not test whether the model knows known vulnerabilities. It delivers a repository with injected flaws and asks the agent to find the bug and write a patch that fixes the problem without breaking existing tests. It is a security engineering test, not CVE memorization.

In this article you will see:

• What CyberBench asks the agent to do.
• How the evaluation works (steps, limits, approval criteria).
• The official ranking from November 5, 2026 and what the numbers show.
• Why this benchmark matters for people who use code agents.

What is CyberBench

CyberBench is a benchmark that evaluates whether an AI agent can find a security flaw in a repository and produce a patch that fixes it. Each task delivers a repository with an injected vulnerability and the existing tests. The agent needs to locate the bug, understand the root cause and write a fix that passes the security tests and the regression tests.

The idea is to measure analysis and correction capability, not pattern detection. The agent has no access to the internet, CVE databases or static analysis tools. It can inspect the code, run the tests and iterate on the fix. This is closer to the real work of a security developer than running an automated scanner.

This differentiates CyberBench from CWE-bench v2, which measures the ability to identify the vulnerability type. Here the agent needs to fix the problem. It also differs from Vals CyberOffense, which measures whether open models can invade corporate networks.

How the evaluation works

Vals AI evaluates agents on 48 public CyberBench tasks. Each task provides a repository with an injected vulnerability. The agent receives a bash terminal, an offline sandbox and the benchmark's standard prompt. It can inspect the code, run the tests and iterate on the fix, but it cannot use the internet, consult CVE databases or use external static analysis tools.

The submission contains the patch produced by the agent. Vals applies the patch and runs the security tests and regression tests in a new container. The agent has up to 1,000 steps, a 6 hour wall clock limit and a 180 second timeout per action. Tool output is truncated to 10,000 characters at the edges.

A patch passes if it fixes the vulnerability and does not break existing tests. There is no partial credit for explanation: either the fix works, or it does not.

The official ranking from November 5, 2026

Vals tested 48 agents on CyberBench. Wonder, from Meta, leads with 80.36% accuracy, followed by Claude Opus 5.5 with 75.00% and GPT-6 Astra with 72.22%. Most agents are below 60%.

The difference between first and second is more than 5 percentage points. This shows that finding and fixing security flaws is still a difficult task, even for the strongest models.

Vals also compares agents on cost and latency per task. Wonder is cheaper per attempt, but Claude Opus 5.5 is the second most accurate.

How to test today

For anyone who wants to try it out, Vals makes the methodology and results available on the official leaderboard. You can run the same agents on your own repository and compare the result with the published ranking.

Vals also publishes the evaluation code and the prompts used, which makes it possible to reproduce the tests on other models and environments. This is useful for teams that want to validate whether a new agent meets their security needs before adopting it in production.

Why this benchmark matters

The CyberBench measures a capability that correctness benchmarks do not capture: the ability to find and fix security flaws in real code. This is useful in three real world scenarios.

First, automated security review: when you want an agent to review code before it goes to production. Second, vulnerability triage: when you have many alerts and need to prioritize which to fix. Third, validation of code agents that need to operate in repositories with security requirements.

The 80.36% result is also a reminder of the limits. Even the best agent is wrong in almost 20% of cases. For those who use code agents day to day, this reinforces the need for human review and testing, as discussed in the analysis of Vals CyberOffense in security.