agent-guardian 1.0.0rc11

Open-source red teaming toolkit for AI agents, RAG systems, MCP servers, and tool-using LLM applications.

AgentGuardian

Red-team your AI agents before attackers do.

Docs · Quickstart · Try the demo agent · Attack library · CI/CD · Sample report

---

AgentGuardian is an open-source red-teaming toolkit for AI agents. It scans your agent, maps the attack surface, runs the relevant adversarial agents, and generates evidence-backed findings for you to review — and fix the vulnerabilities before they reach production.

▶ Watch the demo to see how AgentGuardian finds vulnerabilities in a live scan.

Getting started

1. Install

pip install agent-guardian

or

uv tool install agent-guardian

2. Configure a model provider

AgentGuardian drives its attacks with an LLM. Export a key for your provider — Gemini, OpenAI, or Anthropic:

export GEMINI_API_KEY=...        # or OPENAI_API_KEY / ANTHROPIC_API_KEY

For every supported provider and the full set of configuration options, see the configuration guide.

3. Scan an agent

No agent of your own yet? Point it at the hosted demo target — a deliberately vulnerable "finbot" banking agent:

agent-guardian scan \
  --endpoint https://agent-guardian-testbench-u6tm6gzysq-uc.a.run.app/finbot/chat \
  --model gemini:gemini-3.5-flash \
  --mode fast \
  --output pdf --output-path report.pdf

To scan your own agent instead, swap --endpoint for any target — a hosted URL, a --system-prompt file, or a --framework object (see What you can scan).

4. Review the findings

AgentGuardian opens a live dashboard while it runs (http://127.0.0.1:7474) and writes an evidence bundle — findings, transcripts, and your PDF report — under ~/.agentguardian/scans/<scan-id>/.

What you can scan

Scan an HTTP agent

agent-guardian scan \
  --endpoint http://localhost:8000/chat \
  --model gemini:gemini-3.5-flash \
  --mode smart

Scan a system prompt

agent-guardian scan \
  --system-prompt ./prompts/customer-support-agent.txt \
  --model gemini:gemini-3.5-flash \
  --mode fast

Scan a LangGraph agent

agent-guardian scan \
  --framework langgraph \
  --framework-ref my_app.graph:graph \
  --model gemini:gemini-3.5-flash \
  --mode smart

Scan a Google ADK agent

agent-guardian scan \
  --framework adk \
  --framework-ref my_agent.agent:root_agent \
  --model gemini:gemini-3.5-flash \
  --mode smart

For more targets (MCP, OpenAPI, WebSocket, browser flows), see the target adapters guide.

What AgentGuardian catches

AgentGuardian tests agentic risks that normal prompt scanners miss:

• Prompt injection and goal hijack
• Unsafe tool calls and tool chaining
• Privilege abuse
• RAG poisoning and indirect prompt injection
• Memory and context poisoning
• Sensitive data leakage
• Agent-to-agent manipulation
• Cascading failures
• Trust exploitation and unsafe outputs
• Goal drift and untraceable behavior

Reports and evidence

Every scan writes a local evidence bundle under ~/.agentguardian/scans/<scan-id>/:

• scan.json — machine-readable findings, signed (HMAC-SHA256 + Ed25519)
• events.jsonl — the scan timeline
• probe/ — per-probe requests, responses, verdicts, and evidence
• forensic_manifest.json — integrity manifest for the bundle
• a live local dashboard — browse findings, transcripts, and exports

Generate shareable or CI-ready reports in any format on demand:

agent-guardian report SCAN_ID --output sarif --output-path scan.sarif   # GitHub Security
agent-guardian report SCAN_ID --output md                                # Markdown
agent-guardian report SCAN_ID --output pdf  --output-path report.pdf      # shareable PDF

Formats: json · sarif · junit · md · gitlab · pdf. Stored evidence can be verified with agent-guardian verify.

How it works

Every scan follows the same loop:

Target → surface mapping → adversarial agents → AIVSS-scored findings → evidence bundle

For the full workflow, see how AgentGuardian works.

Scan modes

• fast — quick local feedback
• smart — broader coverage for development and pull requests
• full — release gates and audit evidence

Use full when you need AIVSS-scored findings for CI/CD gates.

Commands

Command	What it does
agent-guardian scan	Run an adversarial swarm scan against a target
agent-guardian report <id> --output FMT	Regenerate a report — json · sarif · junit · md · gitlab · pdf · badge
agent-guardian gate <id> --fail-under N	Apply pass/fail thresholds to a stored scan (CI exit codes)
agent-guardian serve	Start the local dashboard
agent-guardian scans list / delete	List or delete stored scans (delete --older-than 30d for bulk cleanup)
agent-guardian config show / init	Inspect the effective config / scaffold a config file
agent-guardian verify <report>	Verify the HMAC-SHA256 + Ed25519 signatures on a report
agent-guardian last-score	Print the AIVSS of the most recent scan
agent-guardian doctor	Verify the install, provider keys, and prerequisites
agent-guardian telemetry status	Manage opt-in telemetry (enable / disable)
agent-guardian version	Print the installed version

Run any command with --help for its full options, or see the CLI reference.

CI/CD with GitHub Actions

The shipped composite action runs a scan, uploads SARIF to GitHub Code Scanning, and (optionally) posts a summary comment on the pull request:

name: AgentGuardian

on:
  pull_request:
  push:
    branches: [main]

jobs:
  red-team:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write   # upload SARIF to Code Scanning
      pull-requests: write     # post the summary comment
    steps:
      - uses: actions/checkout@v4

      - uses: glacien-technologies/agent-guardian/.github/actions/agentguardian-scan@v1
        with:
          endpoint: http://localhost:8000/chat
          model: gemini:gemini-3.5-flash
          mode: full
          fail-under: "80"
          max-critical: "0"
          comment: "true"
        env:
          GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}

The job fails when the gate (fail-under / max-critical) is breached. For GitLab, Bitbucket, raw-CLI, and fleet/nightly setups, see the CI/CD guides — including the parallel suites guide for scanning many agents from one file.

Standards and coverage

AgentGuardian maps its shipped probes to:

• OWASP Top 10 for Agentic Applications
• MITRE ATLAS
• CSA Agentic AI Red Teaming Guide

The exact agents and probes that ran against your target are enumerated in every scan report (coverage in scan.json). The full probe-to-standard mapping lives in the OWASP mapping and the framework coverage matrix.

Privacy & telemetry

Telemetry is opt-in and disabled by default. Out of the box AgentGuardian sends nothing — no analytics ping, no install ping, no scan counts. Telemetry only activates after you explicitly opt in. Once enabled, it sends anonymous operational counts (agents dispatched, attempts, findings) plus a locally generated, anonymous install id (a random UUID stored at ~/.agentguardian/install_id, with no link to your identity).

Manage it any time:

agent-guardian telemetry status     # show current state
agent-guardian telemetry enable      # opt in
agent-guardian telemetry disable     # opt out

AgentGuardian never collects prompts, agent responses, target URLs, headers, secrets, API keys, transcripts, reports, evidence files, tool inputs or outputs, or customer data.

Run from source

To run AgentGuardian from a source checkout instead of the published package:

# clone
git clone https://github.com/glacien-technologies/agent-guardian.git
cd agent-guardian

# virtual environment
python3.11 -m venv .venv
source .venv/bin/activate

# install the checkout in editable mode
pip install -e ".[dev]"

# run it from source
agent-guardian doctor
agent-guardian scan \
  --endpoint http://localhost:8000/chat \
  --model gemini:gemini-3.5-flash \
  --mode fast

For contribution guidelines, see the contribution guide.

Contributing

We welcome new probes, new adapters, and new attacker logic. Start with the contribution guide and the good first issue label.

All commits must be DCO-signed:

git commit -s

By participating you agree to CODE_OF_CONDUCT.md and the ethics policy. AgentGuardian is for testing systems you own or are explicitly authorised to test.

Community

Join us on Discord for quickstart help, probe design, adapter questions, and roadmap discussion. For longer-form support channels, see the support guide.

Security

To report a vulnerability, see SECURITY.md. Do not open public issues for security reports.

License

Apache-2.0. See LICENSE and NOTICE.

AgentGuardian is a trademark of Glacien Technologies. See TRADEMARKS.md for usage guidelines.