agent-guardian 1.0.0rc1

Adversarial swarm framework for agentic AI red-teaming. Eleven specialist agents, deterministic AIVSS score, OWASP ASI / MITRE ATLAS / CSA aligned.

AgentGuardian Open

The first open-source adversarial-swarm framework for agentic AI red-teaming. Eleven specialist agents attack your AI agent in parallel under a Swarm Commander LLM. Output: a deterministic 0–100 AIVSS score aligned with the OWASP Top 10 for Agentic Applications 2026, MITRE ATLAS v5.4.0, and the CSA Agentic AI Red Teaming Guide.

Why

Single-chain red-teaming tools send one prompt at a time. Production agents compose tools, hold memory, talk to other agents, and run real code — and that surface needs eleven attackers working in concert.

AgentGuardian deploys a swarm: a reconnaissance agent maps your target, then ten specialist agents (one per OWASP ASI category) attack concurrently, coordinated by a Swarm Commander that re-tasks idle agents and stops early on convergence. Every finding is triple-tagged with OWASP ASI, MITRE ATLAS, and CSA Agentic-RT categories.

Read the full rationale: Why we built this.

How it compares

Tool	Multi-agent swarm	Agentic-AI focus	Standards alignment	Open formula	License
PyRIT	no	no	NIST AI RMF (partial)	no	MIT
garak	no	no	own taxonomy	no	Apache-2.0
Promptfoo	no	no	own taxonomy	no	MIT
Inspect	no	no	own taxonomy	no	MIT
DeepTeam	no	no	OWASP LLM Top 10	no	Apache-2.0
AgentGuardian	yes	yes	OWASP ASI + ATLAS + CSA + AIVSS	yes	Apache-2.0

Quickstart

pip install agent-guardian

# Pick an LLM backend, or use --model stub for zero-key testing.
export OPENAI_API_KEY=sk-...

# Scan a system prompt
echo "You are a helpful customer-support bot." > prompt.txt
agent-guardian scan --system-prompt prompt.txt

# Live dashboard at http://localhost:7474
agent-guardian serve

# Marketing badge
agent-guardian badge $(agent-guardian last-score) --svg > badge.svg

Full walkthrough: Five-minute quickstart.

Run with Docker

docker build -t agent-guardian:dev .
docker run --rm -p 7474:7474 agent-guardian:dev serve --host 0.0.0.0

Or with the bundled compose file:

docker compose up --build

Architecture

                          ┌─────────────────────────────┐
                          │     Swarm Commander LLM     │
                          │  (orchestration & dispatch) │
                          └──────────┬──────────────────┘
                                     │
                ┌────────────────────┼────────────────────┐
                │                    │                    │
                ▼                    ▼                    ▼
        ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
        │  Recon Agent │     │   Shared     │     │   Findings   │
        │  (map target)│◀───▶│ Vector Memory│◀───▶│  Aggregator  │
        └──────────────┘     └──────────────┘     └──────────────┘
                                     ▲
                ┌────────────────────┼────────────────────┐
                │   Ten ASI-aligned specialist attackers  │
                │   running in parallel                   │
                │                                         │
                │   ASI01  Goal Hijack                    │
                │   ASI02  Tool Misuse                    │
                │   ASI03  Privilege Abuse                │
                │   ASI04  Supply Chain                   │
                │   ASI05  Unauthorised Code Execution    │
                │   ASI06  Memory Poisoning               │
                │   ASI07  Agent-to-Agent Compromise      │
                │   ASI08  Cascading Failures             │
                │   ASI09  Trust Exploitation             │
                │   ASI10  Rogue Agent / Drift            │
                └─────────────────────────────────────────┘

Full architecture: docs/architecture.

Status

Active development, pre-1.0. The swarm, the scorer, the dashboard, and the signed-report pipeline are all in place. v1.0 ships on PyPI at M15.

Roadmap: docs/roadmap.

Documentation

• Why we built this
• Quickstart
• Architecture
• AIVSS formula
• Adapters
• API reference
• Ethics and responsible use
• Roadmap

Contributing

We welcome probes, adapters, bug reports, and PRs. See CONTRIBUTING.md. All contributions require a DCO sign-off.

Security

See SECURITY.md for responsible-disclosure policy. If you find a vulnerability in AgentGuardian itself, please email [e-mail] instead of filing a public issue.

Ethics

AgentGuardian Open is for testing systems you own or are explicitly authorised to test. Use against third-party systems without authorisation is unlawful in most jurisdictions and a violation of these terms. See Ethics.

License

Apache License 2.0. See LICENSE and NOTICE.

Trademark

"AgentGuardian" is a trademark of Glacien Pte. Ltd. See TRADEMARKS.md.