VulnHunter: Capital One’s Open-Source AI Security Tool

Capital One open-sourced VulnHunter on July 17 — an agentic AI security scanner built to reason through your code the way an attacker would, not the way a pattern-matcher does. It runs as a Claude Code skill, requires Claude Opus, and ships under Apache 2.0. The target is the single biggest problem with every SAST tool in existence: a false positive rate that trains developers to ignore security alerts entirely.

SAST Is Broken and AI Made It Worse

Traditional static analysis tools produce false positive rates between 60 and 90 percent. Untuned, some hit 91 percent noise. That number was already a governance problem — now it’s a crisis. Cloud Security Alliance data shows that in codebases with active AI-assisted development, monthly SAST findings have jumped tenfold. Scanners built for human-paced code review were never designed to absorb that volume, and developers have responded the only rational way: they stopped reading alerts.

Capital One built VulnHunter against exactly this backdrop. As they put it: “The defensive tools to address this reality need to be just as widely distributed, tested, and improved as the codebases they protect.” The logic is sound — if AI is generating the code, the security tooling has to match that velocity.

Attacker-First, Not Sink-First

Every mainstream SAST scanner works backward from dangerous function calls — the “sinks” — searching for code paths that could reach them. VulnHunter inverts this. It starts at real attacker entry points: APIs, network messages, file uploads, form inputs. From there it traces forward through application logic and security checkpoints, following the actual path an adversary would take.

The difference matters. Sink-first analysis flags patterns that could be exploitable in some configuration. Forward analysis determines whether an attacker can actually get there. It’s the difference between “this function is dangerous” and “an attacker can reach this function with malicious input after bypassing your authentication layer.”

The Falsification Engine

This is the part worth paying attention to. After the scanner surfaces a potential vulnerability, it doesn’t report it immediately. Instead, VulnHunter runs a separate reasoning workflow specifically designed to disprove its own finding. It looks for logical gaps in the exploit path, unsupported assumptions, and conditions that would block the attack from succeeding. Only if it cannot talk itself out of the finding does the vulnerability get reported.

No production SAST tool does this. The design philosophy for existing scanners is to flag anything suspicious and let developers sort it out — hence the 60-90% noise. VulnHunter’s falsification engine is a direct architectural response to alert fatigue: produce fewer findings, but guarantee they are real.

Hunt, Fix, Verify

VulnHunter is not just a scanner. It is a three-phase remediation pipeline:

• Hunt — Scans entry points, traces exploit paths, applies falsification, surfaces confirmed vulnerabilities
• Fix — Writes security tests and implements targeted code changes in a test-driven workflow
• Verify — A separate, independent agent validates that the fix actually closed the vulnerability

The verify phase runs as its own standalone agent, not a post-processing step in the same run. That separation matters for CI/CD pipelines where you need an isolated check confirming the remediation did not introduce new attack surface.

Getting Started

Installation is straightforward if you already have Claude Code set up:

git clone https://github.com/capitalone/vulnhunter.git
cd vulnhunter
./install.sh

The install script copies skill files to ~/.claude/skills/. Then invoke each phase via the Claude CLI:

# Scan
claude --model opus --add-dir ~/.claude/skills/vulnhunt

# Fix
claude --model opus --add-dir ~/.claude/skills/vulnhunter-fix

# Verify
claude --model opus --add-dir ~/.claude/skills/vulnhunt-fix-verify

The harness ships with synthetic test cases mapped to OWASP NodeGoat, Juice Shop, and WebGoat — validate against known-vulnerable apps before pointing it at your own codebase. A headless agent (vulnhunter-agent/) handles CI/CD integration without interactive prompts.

The Cost Question

VulnHunter runs on Claude Opus, not a cheaper tier. Teams already paying for Opus access through Claude Code subscriptions absorb this naturally. Teams running it purely on API credits should estimate costs against codebase size before wiring it into every PR pipeline.

The economics shift quickly, though. If a single Opus scan eliminates six false positive triage sessions per week — each costing 20-30 minutes of a senior developer’s time — the cost-benefit math resolves fast. The real question is not whether Opus is expensive; it is whether cheaper, noisier alternatives actually cost less when you account for the human time they consume.

Why Now

Seventy-eight percent of Fortune 500 companies now have AI-assisted development in production. That is a lot of code being written by tools that were not designed with security reasoning in mind. VulnHunter is the first serious open-source attempt to close that loop with an auditable tool that any team can inspect, fork, and extend. The Capital One announcement frames this as shared infrastructure — security tooling that improves as more teams use and contribute to it. That framing is correct, and the timing is right. Check the VentureBeat coverage for additional context on Capital One’s open-source strategy behind the release.