AI agents often run shell commands, evaluate code, and fetch URLs. A lot of the glue code around that still has classic issues: secrets in source, unsafe command construction, eval/pickle, SSRF-style fetches, and path handling that can leave a workspace.

I built VulnPath-AI to catch those patterns locally.

What it is

An offline-first Python hybrid scanner aimed at AI/agent application code:

  • Hardcoded secrets
  • Command-injection patterns
  • eval / exec / unsafe pickle
  • SSRF-style URL fetches
  • Workspace path escape (naive joins vs resolve + boundary checks)

It adds path-style context where possible and fix suggestions only — no exploit or attack features.

Why offline

I wanted something you can run on a laptop without sending source to a cloud SAST. Stdlib-first, optional YAML rules, sample suite, benchmark, and SARIF/CI support.

Honest scope

This is not a replacement for enterprise SAST (Semgrep, CodeQL, etc.). It's a focused tool for agent/tool glue risks and for learning how these bugs show up in real agent code.

Try it

 bash git clone https://github.com/zenithVeil/vulnpath-ai.git cd vulnpath-ai pip install -e . vulnpath-ai samples/sqli_example.py # or: python analyze_hybrid.py samples/vulnerable/sqli_string_concat.py