Airlock🛡️

AI-powered security scanner for Hugging Face and GitHub repositories. Detects malware, typosquatting, and supply chain attacks before you clone.

Motivation

In May 2026, a fake OpenAI repository on Hugging Face was downloaded 244,000 times. It contained a multi-stage infostealer that bypassed Windows Defender. RepoGuard would have flagged it with a score of < 10 / 100.

What it detects

• Base64 encoded payloads and eval/exec chains
• SSL verification disabled (allows C2 connections)
• PowerShell execution from Python loaders
• Windows Defender exclusion commands
• Privilege escalation and UAC bypass
• Persistence mechanisms (scheduled tasks, registry)
• Data exfiltration (SSH keys, browser data, crypto wallets)
• Self-deleting scripts
• Dead drop resolver patterns (jsonkeeper, pastebin)
• Malicious pickle files — see below
• VM/sandbox evasion techniques
• Typosquatted account names and dependency names
• Dependencies from non-official indexes, direct URLs, and npm install hooks
• New accounts with artificially trending repos

Supply chain

requirements*.txt, pyproject.toml and package.json are parsed (never resolved or installed). The scanner reports packages that come from somewhere other than the official index, packages installed straight from a URL, npm preinstall/postinstall hooks — which run automatically on npm install — and names one edit from a popular package. Typosquat matching uses Damerau-Levenshtein so that transpositions like reqeusts count as a single edit, because that is how squatting actually looks.

Pickle scanning

torch.save() writes pickles, so pytorch_model.bin, .pt and .ckpt files are programs, not data: loading one runs whatever it names. RepoGuard reads the pickle opcode stream with pickletools.genops and reports the callables the file would import — it never unpickles anything.

Weight files are usually enormous and the pickle inside them is not, so the scanner reads a torch archive's central directory over HTTP range requests and pulls only the data.pkl member: roughly 12 KB out of a 90 MB checkpoint. Files too large to retrieve whole are reported as partially scanned rather than silently passed.

A repo that ships pickle weights with no .safetensors equivalent is flagged as informational — safetensors cannot execute code, so publishing only the executable format is worth a look.

Stack

• Backend: Python + FastAPI
• Scanning: regex pattern registry + Python AST + pickle opcode analysis (no code execution — nothing scanned is ever imported, loaded or unpickled)
• AI layer: Claude API for model card analysis — planned, not yet wired
• Frontend: React + Vite
• Database: SQLite (dev) / PostgreSQL (prod)
• Deploy: Railway or Render

Quick start

cd backend
cp .env.example .env # add your API keys
pip install -r requirements.txt
uvicorn app.main:app --reload

cd frontend
npm install
npm run dev

Run tests

cd backend
pytest app/tests/ -v

API

POST /api/scan
{ "repo_url": "https://huggingface.co/owner/repo" }
GET /api/scans?limit=20 # recent scan history, with finding counts
GET /api/scans/{id} # one past scan and the findings it recorded
GET /api/health

Returns a trust score (0-100), trust level (safe/suspicious/dangerous), detailed pattern matches with file locations and line numbers, and plain English recommendations.

Database

Schema changes go through Alembic (backend/migrations/). The app runs alembic upgrade head on startup, so no manual step is needed in development; a database created before migrations existed is stamped at the baseline revision and upgraded from there without losing rows.

alembic revision --autogenerate -m "what changed"# from backend/
alembic upgrade head