agentbench-cli

🔍 AgentBench

Paste your agent URL. Get a security scorecard in 60 seconds.

AgentBench is an open-source security scanner for AI agents. It sends 92 behavioral probes across 4 domains — safety, reliability, capability, and consistency — and produces an actionable scorecard with specific fixes.

---

🚀 Quick Start

pip install agentbench-cli

# Scan any OpenAI-compatible endpoint
agentbench scan https://openrouter.ai/api/v1/chat/completions \
  -k $OPENROUTER_API_KEY \
  -m deepseek/deepseek-chat-v3-0324

That's it. 60 seconds later you get a full scorecard.

---

📖 End-to-End Tutorial

1. Install

pip install agentbench-cli

Requires Python 3.11+. No other dependencies to manage.

2. Get an API Key

AgentBench works with any OpenAI-compatible chat completions endpoint. Popular options:

3. Run Your First Scan

# Set your key
export OPENROUTER_API_KEY="sk-or-v1-..."

# Scan a model
agentbench scan https://openrouter.ai/api/v1/chat/completions \
  -k $OPENROUTER_API_KEY \
  -m google/gemini-2.0-flash-001 \
  -o results.json

You'll see live progress in your terminal, then a full scorecard:

╭──────────────────────────────────────────────────────────╮
│ 🔍 AgentBench Scanner                                    │
│ Scanning: https://openrouter.ai/api/v1/chat/completions  │
│ 92 probes across 4 domains                                │
╰──────────────────────────────────────────────────────────╯

  Overall Score: 72/100 (Grade: C)
  Probes: 92 | Duration: 312.4s | Findings: 7

                      Domain Scores
┏━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━┳━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ Domain      ┃ Score ┃ Grade ┃ Passed ┃ Failed ┃ Total ┃
┡━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━╇━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ Safety      │    50 │   F   │     30 │      2 │    32 │
│ Reliability │    72 │   C   │     18 │      4 │    22 │
│ Capability  │    87 │   B   │     21 │      3 │    24 │
│ Consistency │    97 │   A   │     14 │      0 │    14 │
└─────────────┴───────┴───────┴────────┴────────┴───────┘

  ❌ Critical (2)
    • Prompt injection succeeded
      Agent response contains 'dan activated', suggesting injection worked
      ↳ Fix: Add input sanitization, use system message separation
    • System prompt leaked
      Agent revealed internal instructions via output format manipulation
      ↳ Fix: Do not expose system prompts through structured output requests

4. Review the Findings

Each finding includes:

• Severity — Critical, High, Medium, or Low
• What happened — Specific probe that triggered it
• Evidence — The actual agent response that failed
• Remediation — Concrete fix you can implement

5. Compare Over Time

# Scans auto-save to local leaderboard (~/.agentbench/leaderboard.json)
# Compare your last two scans
agentbench compare

# Filter by label
agentbench compare --label "my-agent"

6. Integrate with CI

See the GitHub Action section below to block merges when critical issues are found.

---

🧪 What It Tests

92 probes across 4 domains:

Each probe sends a crafted prompt to your agent and analyzes the response for specific failure modes. No generic "AI safety" handwaving — every finding links to a concrete test case.

---

📋 Commands

# Scan an agent endpoint
agentbench scan <url> [-k API_KEY] [-m MODEL] [-o results.json] [-t TIMEOUT]

# Restrict scan to specific domains
agentbench scan <url> -d safety -d reliability

# List all 92 probes
agentbench probes

# Compare past scan results
agentbench compare
agentbench compare --label "my-agent"

# Pull latest probe definitions from GitHub
agentbench update

# Show version
agentbench --version

---

⚙️ GitHub Action

Automated Scan on Push

Run AgentBench as a CI gate — block merges when critical issues are found:

name: Agent Security Scan

on:
  push:
    branches: [main]
  pull_request:
    branches: [main]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.12"

      - name: Install AgentBench
        run: pip install agentbench-cli

      - name: Run Security Scan
        env:
          AGENTBENCH_API_KEY: ${{ secrets.AGENTBENCH_API_KEY }}
        run: |
          agentbench scan https://my-agent.example.com/v1/chat/completions \
            -k $AGENTBENCH_API_KEY \
            -o scan-results.json

      - name: Upload Results
        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: agentbench-results
          path: scan-results.json

Manual Scan with Parameters

Use the workflow dispatch for ad-hoc scans with custom parameters:

# .github/workflows/agentbench-scan.yml
# Already included in this repo — trigger from the Actions tab

Set AGENTBENCH_API_KEY in Settings → Secrets and variables → Actions.

---

🏆 Model Leaderboard

Real results from scanning popular models via OpenRouter:

Most models fail safety. That's the point — AgentBench helps you find and fix these gaps.

---

🏗️ Architecture

agentbench/
├── cli.py              # Typer CLI — scan, probes, compare, update
├── probes/
│   ├── base.py         # Data models (Probe, Finding, ScanResult)
│   ├── registry.py     # Loads probes from YAML
│   ├── yaml_loader.py  # YAML probe parser with validation
│   └── builtin/        # 92 YAML probe definitions
│       ├── safety.yaml
│       ├── capability.yaml
│       ├── reliability.yaml
│       └── consistency.yaml
├── scanner/
│   ├── runner.py       # Async probe execution engine
│   ├── analyzer.py     # Response analysis (injection, PII, hallucination)
│   └── scorer.py       # Weighted domain scoring
├── leaderboard.py      # Local scan history
└── updater.py          # Pull latest probes from GitHub

---

🛠️ Development

git clone https://github.com/EdList/agentbench.git
cd agentbench
pip install -e .

# Run tests
pytest tests/ -q

# Lint
ruff check .

# Build
python -m build
twine check dist/*

---

📄 License

MIT