VulnHunter reviews source code from an attacker's perspective

What this is about

VulnHunter is an open-source tool from Capital One that examines source code from a potential attacker's perspective. The project, created in July 2026, connects three stages: finding vulnerabilities, creating fixes, and independently checking those fixes. According to its documentation, it starts at reachable inputs such as APIs, file uploads, or network messages rather than only flagging suspicious code patterns.

That is relevant to teams sorting many alerts from traditional SAST systems. VulnHunter does not provide automatic truth, however. The workflow is optimized for Claude Code and Claude Opus, and people must review its output.

What VulnHunter actually does

The repository contains three Claude Code skills. /vulnhunt searches from reachable inputs toward dangerous functions. A challenge process then tries to disprove its own finding. Only when the proposed attack chain survives does the tool produce an evidence-based report and a suggested change.

/vulnhunter-fix first expresses a reproducible exploit as a failing security test, changes the code, and tests again. /vulnhunt-fix-verify runs separately with read-only access. It is intended to confirm that the specific attack no longer works after the change. A headless agent and local batch harness support larger workloads.

Why it matters

Traditional scanners often report patterns that look dangerous in theory but cannot be reached in practice. VulnHunter therefore uses forward analysis from a possible entry point to harmful impact. Security teams can use this to prioritize auditable attack paths and connect repairs to reproducible tests.

The Apache 2.0 repository can be inspected and adapted. There is also a clear dependency: the intended setup requires Claude Code and access to Claude Opus. Model charges, runtime, and possible source-code transfer to a model provider belong in every adoption calculation.

In plain language

Imagine inspecting a building. A conventional inspector marks every door with an unusual lock. VulnHunter instead tries to enter through the public entrance, find a complete route to the vault, and then prove that its own route may be blocked. Only if the route remains valid does it suggest a repair and ask for a separate check.

A practical example

A team operates a web application with 40 API endpoints and receives 120 scanner alerts. It runs VulnHunter only against an explicitly authorized test copy. The tool traces a file upload to a library function. The fix skill creates a test that fails before the change, adjusts input validation, and tests again.

A security specialist reviews permissions, data flow, and the patch. Only then does the separate verifier assess whether the reported path has been closed. A vague alert can become an auditable work item, but code review and penetration testing remain necessary.

Scope and limits

• Other models or development environments may work worse or not at all.
• Agentic analysis can draw false conclusions, miss paths, or suggest unsafe changes. Every finding and patch requires human review.
• The tool performs dual-use security work. It must be used only with explicit authorization, and provider safeguards may block requests.
• Confidential source code may leave the workstation. Contracts, retention, telemetry, and privacy must be reviewed first.
• The benchmark harness includes only a small synthetic example, according to the repository. Independent evidence of a generally low false-positive rate is missing.

The best first test is a limited, non-production repository with known vulnerabilities. Teams should record detection rate, false positives, cost, runtime, and test quality.

SEO & GEO keywords

VulnHunter, Capital One, Claude Code, Claude Opus, AI security, source code review, SAST, vulnerability analysis, Apache 2.0, DevSecOps