VulnHunter analyses source code for security defects by simulating an attacker's path rather than matching suspicious patterns. Where a static scanner starts at a dangerous sink and searches backwards for a hypothetical attacker, VulnHunter begins at points an attacker can reach — APIs, network messages, file uploads — and reasons forward to decide whether the path can actually be broken through.
A candidate finding then goes through a falsification workflow built to disprove it, searching for flawed assumptions, logic gaps, and security controls that would block the attack; findings resting on unsupported assumptions are discarded. What survives is emitted with the exploit path mapped, the structural flaw explained, the capabilities or access an attacker would gain listed, and a focused code change proposed for review.
It ships as three Claude Code skills that form a hunt, fix, and verify loop.
| Skill | Phase | Responsibility |
|---|---|---|
| /vulnhunt | Hunt | Maps entry points to dangerous sinks and filters findings through recon, parallel hunt, adversarial disprove, and capability-filter stages, emitting verified issues with an executable exploit and a proposed fix |
| /vulnhunter-fix | Fix | Test-driven remediation: writes an exploit demo, adds a failing security test, implements the fix, confirms the exploit is blocked without regressions, and opens a reviewable pull request |
| /vulnhunt-fix-verify | Verify | A separate read-only agent that independently checks whether a finding was remediated and emits a per-finding verdict |
Two wrappers cover scale. A config-driven headless runtime clones targets, runs the scanner, publishes results, and opens GitHub issues for confirmed bugs through the Anthropic API. A local harness batch-scans a list of repository URLs with resume, status, and collect commands, and a benchmarking mode clones a corpus of known-vulnerable targets, scans, judges with a model, and tallies accuracy metrics against ground-truth files you supply.
The loop depends on multi-step reasoning and is built for Claude Opus running in Claude Code, with model access supplied by the user. The verifier runs under a restricted tool envelope with no shell execution and no network access. Because vulnerability discovery and exploitation are dual-use, running it against an Anthropic account not enrolled in the Cyber Verification Program can have requests blocked by real-time cyber safeguards.
Features
- Forward attacker analysis: analysis starts at attacker-accessible entry points and reasons towards sinks, rather than sink-first
- Falsification engine: a structured pass tries to disprove each finding and drops those depending on unsupported assumptions
- Evidence-backed remediation: surviving findings carry an exploit path, the structural flaw, the attacker capability gained, and a targeted patch
- Closed loop: /vulnhunt, /vulnhunter-fix, and /vulnhunt-fix-verify cover hunt, fix, and independent verification
- Read-only verifier: the verification agent has file and search tools only, with no Bash execution and no network access
- Headless runtime: clones targets, runs scans, publishes results, and files GitHub issues for CI or unattended use
- Batch harness: scans a list of repositories with resume, status, and collect steps for centralised review
- Benchmark mode: clone, scan, model-judge, and tally against a ground-truth corpus you define per repository
