Back to search
vulnhound
MarketplaceWorks with Claude CodeGitHub
vulnhound: token-thrifty vulnerability finder for Python projects
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/bu-ra34--vulnhound)HTML
<a href="/listings/bu-ra34--vulnhound"><img src="/api/listings/bu-ra34--vulnhound/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
As a Claude Code plugin:
/plugin marketplace add bu-ra34/vulnhound /plugin install vulnhound@vulnhound
The CLI itself (the skill needs it on your PATH):
pip install vulnhound
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 15 lines
- {
- "name": "vulnhound",
- "description": "vulnhound: token-thrifty vulnerability finder for Python projects",
- "owner": {
- "name": "bu-ra34"
- },
- "plugins": [
- {
- "name": "vulnhound",
- "source": "./",
- "description": "Token-thrifty vulnerability finder for Python: finds, proves, fixes without reading the whole repo.",
- "version": "0.1.0"
- }
- ]
- }
README.md · README · 67 lines
- # vulnhound
- **Finds real vulnerabilities. Proves them. Fixes them. Without burning your tokens.**
- > Status: pre-alpha. Only `estimate` and `version` work today; the rest is on the roadmap.
- [Türkçe](README.tr.md)
- ## Why
- Asking an LLM to security-review a codebase usually means feeding it the whole repo.
- That is expensive, slow, and most of those tokens are code that has nothing to do with
- any vulnerability. vulnhound avoids this: cheap static analysis finds candidate flows
- first, so the LLM only reads a short list and a few small code slices.
- ## How it works
- - **The core is the eyes.** A stdlib-only Python program builds an AST index and runs
- taint analysis from untrusted sources to dangerous sinks. It never calls an LLM and
- prints short JSON.
- - **Claude is the brain.** Inside Claude Code, the bundled skill tells Claude to run the
- CLI, read only that output plus small slices, judge each finding, write a PoC for the
- likely-real ones, prove it in a sandbox, and patch it.
- Zero runtime dependencies. Python 3.10+. MIT licensed.
- ## Install
- As a Claude Code plugin:
- ```
- /plugin marketplace add bu-ra34/vulnhound
- /plugin install vulnhound@vulnhound
- ```
- The CLI itself (the skill needs it on your PATH):
- ```
- pip install vulnhound
- ```
- ## CLI
- | Command | What it does |
- |---|---|
- | `vulnhound analyze [path] [--json] [--no-cache] [--min-severity S]` | Find candidate vulnerabilities. Exit 1 if any. Findings sharing a sink/helper are grouped (`--no-group` to list all); tests/examples/docs dirs and `test_*.py` are skipped unless `--include-tests`; `--include-local-sources` adds local/CLI inputs as sources. |
- | `vulnhound slice <id>` | Minimal code slice for one finding. |
- | `vulnhound prove <id> --poc FILE [--root DIR] [--backend auto\|docker\|unshare] [--timeout S] [--json] [--verbose]` | Run a PoC in a sandbox (no network, resource-limited). Prints PROVEN (exit 0), SAFE (1) or UNVERIFIED (3). |
- | `vulnhound estimate [path] [--json]` | Rough token cost (bytes/4) of reading the whole repo. |
- | `vulnhound report [--format json\|sarif\|markdown] [-o FILE]` | Render findings. |
- | `vulnhound version` | Print version. |
- Exit codes: 0 ok / no findings, 1 findings found, 2 usage error (e.g. unknown id), 3 UNVERIFIED (prove).
- ## Proving a finding
- `vulnhound prove <id> --poc poc.py` runs a small standalone Python PoC in a sandbox (network off, memory/CPU/time limited, project mounted read-only) and reports a verdict. The PoC triggers the vulnerable code with a malicious input and calls `proven("detail")` from `vulnhound.poc` when the exploit works, or `safe("detail")` when it is blocked. `canary_path()` returns a secret file outside the project, for path-traversal PoCs to read. No marker and exit 0 means SAFE; sandbox missing, timeout or a crashing PoC means UNVERIFIED. The PoC never runs on the host.
- Caveat: the sandbox needs docker or unshare. The PoC runs against your project, so third-party imports must be importable inside it: the `unshare` backend sees the host venv read-only, while the `docker` backend uses a clean python image, so PoCs needing framework imports may report UNVERIFIED there. Pick `--backend unshare` when the PoC imports your installed deps.
- ## Security
- Scan only code you own. See [SECURITY.md](SECURITY.md).
- ## License
- MIT
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.