A scanner that refuses to believe its own findings

What it does

VulnHunter is an open-source agentic security scanner, originally built inside Capital One and now carried forward by a community fork. Instead of the usual sink-first pattern matching that floods you with false positives, it starts at attacker-accessible entry points (APIs, file uploads, network messages) and reasons forward along the actual attack path. Findings that survive get an executable proof-of-concept and a proposed fix; the loop closes with a separate read-only verifier agent that independently judges whether the fix worked.

The interesting bit

The falsification engine. After a candidate vulnerability is found, a structured workflow tries to disprove it — hunting for flawed assumptions, logic gaps, or controls that block the attack. Per the fork’s own benchmarks, 55% of candidate findings get eliminated or downgraded before reaching you. The fork’s honest twist: six scans of the same commit produced a 14× spread in confirmed findings, which the maintainer treats as the first vulnerability to fix — sandboxed, containerized exploit validation is the in-progress answer.

Key highlights

• Three composable skills: /vulnhunt (hunt), /vulnhunter-fix (test-first RED/GREEN remediation with a PR), /vulnhunt-fix-verify (independent read-only verification).
• Harness-portable: skills are plain prompt files with an explicit environment contract (VULNHUNT_SKILLS_DIR, VULNHUNT_MODEL, etc.), not bound to Claude Code like upstream.
• vulnhunter-run operator for unattended nightly runs — clone, hunt, manifest — with explicit stop rules.
• Benchmark numbers from six scans of one production Go service: 42/42 confirmed findings carried passing executable exploit tests; 315 attacker-controlled inputs inventoried and traced in a single scan.
• You supply your own model access; the methodology is prompt procedure, not tool binding.

Caveats

• Sandbox-first exploit validation and measured-impact PoCs are explicitly “in progress” — some exploit tests may still be static traces.
• The dual-use nature means commercial models may refuse or rate-limit aggressive exploitation; the project was developed and tested on open-weight, uncensored models, which is a deployment decision in itself.
• The headline benchmark numbers come from the maintainer’s own runs against one service; raw artifacts are retained by them, not published.

Verdict

Worth a look if you run an agent harness and want SAST findings that arrive pre-litigated with PoCs instead of pattern-match noise. Skip it if you need a finished, vendor-neutral product today — parts of the hardening roadmap are still being built.