agentvet

AgentVet

AI Agent Security Scanner — detect prompt injection, tool auth bypass, and data leaks before attackers do.

One command. Find vulnerabilities in your AI agent code.

---

Features

• 4-Tier Detection Pipeline
  • L1 — Regex + AST patterns (fast, broad coverage)
  • L2 — Ollama semantic filter (removes false positives, ~24s/batch)
  • L3 — DeepSeek deep audit (attack paths + PoC + CVSS + fix, on HIGH/CRITICAL only)
  • L4 — Cross-finding attack chain synthesis (how vulns combine into full attack campaigns) — unique to AgentVet
• 22 Detection Rules across 7 categories: prompt injection, tool authorization, data leakage, framework security, secret exposure, MCP config, and supply-chain
• Web Dashboard — React + Tailwind UI with scan history and per-finding drilldown
• CLI Tool — agentvet scan ./my-agent --depth 3
• Zero data leaves your machine (L1 is local-only; L2 uses local Ollama; L3 is opt-in via API key)

---

Quick Start

Prerequisites

• Python 3.10+
• Ollama (optional, for L2 semantic filtering)
  • Pull a model: ollama pull qwen3:8b
• DeepSeek API key (optional, for L3 deep audit)

Install

git clone https://github.com/tlyyxjz/agentvet.git
cd agentvet

# Using pip
pip install -e .

# Or using uv
uv sync

Usage

CLI:

# L1 only (fastest, some noise)
agentvet scan ./my-agent --depth 1

# L1 + L2 (removes false positives)
agentvet scan ./my-agent --depth 2

# Full pipeline (L1 + L2 + L3 deep audit)
agentvet scan ./my-agent --depth 3

# JSON output
python quick_scan.py --target ./my-agent --depth 3 --json

Web Dashboard:

# Start backend
uvicorn web.main:app --host 0.0.0.0 --port 8765

# Start frontend (in another terminal)
cd frontend && npm install && npm run dev

Open http://localhost:5173 to use the web dashboard.

---

Configuration

Copy .env.example to .env and set your values:

---

Architecture

scan target
  │
  ├─ L1: RegexRule + ASTRule  (~1s,  free, ~60% coverage)
  │   └─ 17 rules auto-discovered from scanner/rules/
  │
  ├─ L2: Ollama qwen3:8b     (~24s, free, removes ~30% noise)
  │   └─ Batch classify: REAL vs NOISE
  │
  ├─ L3: DeepSeek-chat       (~10s/finding, ~$0.01/scan)
  │   └─ Per-finding: attack path + exploit demo + CVSS + fix
  │
  └─ L4: DeepSeek-chat       (~15s, one call total)
      └─ Cross-finding chain synthesis: how vulns combine into attack campaigns

Detection Rules

All 22 rules auto-discover from scanner/rules/ — no registration needed. Each rule maps to OWASP LLM Top 10 (2025) and OWASP Agentic AI Top 10 categories.

Project Structure

agentvet/
├── scanner/          # Core scan engine
│   ├── engine.py     # ScanEngine + Rule base classes
│   ├── findings.py   # Finding + ScanReport data models
│   ├── l2_filter.py  # L2 Ollama semantic filter
│   ├── l3_audit.py   # L3 DeepSeek deep audit
│   └── rules/        # Detection rules
├── cli/              # CLI entrypoint
├── web/              # FastAPI backend
├── frontend/         # React + Vite + Tailwind
├── docs/             # Documentation
└── tests/            # Test suite (coming soon)

---

Roadmap

• npm distribution (npx agentvet scan)
• More detection rules (CSRF, SSRF, path traversal in agent tools)
• Docker image
• Supabase migration for cloud deployment
• VS Code extension
• GitHub Action (agentvet/scan)

---

License

MIT — see LICENSE.