VulnHunter: Capital One's agentic AI code security tool

TL;DR Highlight

Capital One open-sourced a security agent that tries to disprove its own findings to cut false positives.

Who Should Read

A security-minded backend developer who spends too much time triaging false positives from SAST tools.

Core Mechanics

Traces forward from entry points
- The post says conventional tools work sink-first, searching backward from dangerous code for a hypothetical attacker.
- VulnHunter starts at APIs, network messages, or file uploads and reasons forward through the logic.
It tries to disprove its own findings
- After finding a vulnerability, it runs a reasoning step (falsification engine) built to break its own argument.
- It looks for broken assumptions and gaps in the exploit path, and discards unsupported findings.
Only survivors get a fix
- It gathers evidence across the codebase to map the full exploit path.
- It explains the defect, the attacker's gained access, and proposes code changes for review.
Needs Claude Opus 4.8 and Claude Code
- It is optimized for Claude Opus 4.8 and first implemented as a Claude Code skill.
- The post only says the framework may work with other harnesses and models.

Community Reactions

Opinion: it's a methodology, not a tool
- One commenter says it's mostly markdown files describing how to run an assessment.
- They add that security consultancies already use such methodologies, so putting them into an LLM makes sense.
Pushback: not truly open source
- Commenters say it needs Claude Code CLI and Opus access, so it isn't really open source.
- One commenter applied to Anthropic's verification portal weeks ago and never heard back.
Opinion: no moat, and false security worry
- A commenter says these security harnesses all look alike, so there is no moat.
- Links to similar tools from Visa and Cloudflare were shared.
- Another commenter says such tools find real bugs but worries about a false sense of security.
Opinion: find-and-falsify is interesting
- One commenter found it interesting that the same model both finds and falsifies findings.
- Others asked Capital One to share more detail on internal use.

How to Apply

Run the falsification prompt on one module today
- Pick one API handler under security review and paste the prompt below into Claude.
- Check only the findings that survive the disproof step.
Add a disproof step to existing scanner output
- Take 10 items from your SAST false-positive list and trace them forward from entry points.
- Mark cases where the attack path breaks as likely false positives.
Pilot the Quickstart on one internal repo
- First confirm you have access to Claude Opus and Claude Code.
- Run it on a small repo and compare results with your manual triage.

Try It

copy & run

# Paste into Claude or Claude Code (simplified version of the VulnHunter approach)

You are a security reviewer. Find vulnerabilities in the code below that an external attacker could exploit.
Step 1: List attacker-reachable entry points (APIs, network messages, file uploads).
Step 2: From each entry point, trace forward through data transformations and security checks, and list candidate vulnerabilities.
Step 3: Try to disprove each candidate. Look for unsupported assumptions, gaps in the exploit path, and conditions that block the attack. Discard candidates that rely on unsupported assumptions.
Step 4: For survivors only, give a defect explanation, the access the attacker gains, and a minimal code fix.

Code:
{{code}}

Next step: also print the discarded candidates and why, to check whether the disproof step is sound.

Terminology

agenticInstead of answering once, the model plans steps, reads code, and works on its own. It is like a security engineer tracing code by hand.

false positiveAn alert for something that isn't actually a problem. Too many of them make developers ignore alerts.

sink-first analysisStarts at dangerous code like risky function calls and works backward to see if an attacker could reach it. It reasons from destination back to source.

falsification engineA step that tries to prove its own findings wrong. It is like appointing a devil's advocate in code review.

Claude Code skillA bundle of instructions and tools Claude Code follows for a specific task. Think of it as an operating manual for the agent.