Capital One Open-Sources VulnHunter, an AI Tool That Finds Code Vulnerabilities Before Attackers Do

Quick Facts

• Capital One released VulnHunter on GitHub under an Apache 2.0 license, built on Anthropic’s Claude Opus and Claude Code.
• The tool scans from attacker entry points forward, then runs a falsification engine that eliminates roughly half of all candidate findings before flagging them.
• Capital One tested VulnHunter internally across thousands of repositories spanning tens of business areas before releasing it publicly.

Capital One released VulnHunter on Thursday, an open-source AI security tool that maps how attackers would reach exploitable code and proposes targeted fixes before any vulnerable line ships to production. The tool is available now on GitHub under an Apache 2.0 license.

VulnHunter requires access to Claude Opus 4.8 and a working Claude Code environment. The repository includes a Quickstart guide, architecture documentation, and annotated example workflows.

How It Works

Most vulnerability scanners start at dangerous code patterns and work backward. VulnHunter flips that model. It starts at attacker-accessible entry points such as APIs, network messages, and file uploads, then reasons forward to determine whether an attacker can actually break through.

After surfacing a finding, a falsification engine runs a structured workflow designed to disprove its own conclusions. It searches for unsupported assumptions, logical gaps in the exploit path, and conditions that would block the attack. Any finding that relies on shaky assumptions is discarded. About half of all candidate findings are eliminated at this stage.

Findings that survive then receive a full remediation package. VulnHunter maps the exact exploit path, explains the structural flaw, details what access an attacker would gain, and generates targeted code changes. It writes an exploit demo, creates a failing security test, implements a fix, verifies the exploit is blocked without regressions, and opens a reviewable pull request.

Architecture and Model Requirements

VulnHunter ships as three composable Claude Code skills that form a complete automated remediation loop. The pipeline runs four stages: Recon, Parallel Hunt, Adversarial Disprove, and Capability Filter. Only verified issues with an executable exploit and a proposed fix reach a developer.

The tool was built for Claude Opus. Its low false-positive discipline depends on frontier-class reasoning, though Capital One notes the underlying orchestration patterns can be adapted to other advanced foundation models.

Users running VulnHunter against an Anthropic account not enrolled in Anthropic’s Cyber Verification Program may have requests blocked by real-time safeguards, and usage may be flagged for cyber abuse. Capital One includes this dual-use warning prominently in the repository documentation.

Internal Results and Company Context

Capital One ran VulnHunter on its own codebase before the public release. The company identified and remediated vulnerabilities across thousands of repositories spanning tens of business areas. Work that previously required significant time and manual triage now produces verified, actionable findings at a fraction of the cost in time.

The release carries weight given Capital One’s history. A 2019 breach compromised personal data for roughly 106 million people across the United States and Canada and resulted in an $80 million federal fine. The attacker exploited a Server-Side Request Forgery vulnerability in a misconfigured web application firewall.

Capital One is the only major U.S. bank to migrate entirely to the public cloud. The company reported $669 billion in total assets as of December 31, 2025.

“At Capital One, we decided that the right response to AI-enabled threats wasn’t to wait, but to build cutting-edge AI-driven defenses and put them in the hands of defenders everywhere,” the company said in a statement announcing the release.

What This Means for Security Teams

A major financial institution open-sourcing an offensive security tool sets a precedent. Teams that adopt VulnHunter get a pre-production vulnerability scanner that eliminates most false positives before they reach a developer’s queue and delivers fixes alongside findings.

The catch is the model dependency. Organizations without access to Claude Opus 4.8 or those not enrolled in Anthropic’s Cyber Verification Program will face friction before they can run the tool at full capacity.

Read more: Capital One releases VulnHunter, an open-source AI tool that finds software flaws before hackers do