Get your AI agents EU AI Act compliant in 30 minutes.
Automated red-teaming for AI agents. 130+ attack patterns, 8 categories, behavioral baselining, and a quantifiable Security Score — all in one CLI, API, and CI-friendly platform.
Your AI agents are running. Your security testing isn't.
Web-app scanners don't speak LLM. General red-team tools don't understand tool use, MCP, or multi-agent pipelines. AgentGuard was built for this gap.
How AgentGuard works
One CLI. One API. Five things it does better than anything else.
130+ attack patterns
Behavioral baselining
Security Score
EU AI Act reports
Black-box & gray-box modes
One number your CISO, your auditor, and your marketing team can all quote.
AgentGuard's Security Score combines attack coverage, severity weighting, and exploit success rate into a single 0–100 figure. Embed the live badge on your product site the same way you'd embed an SSL rating — only this one actually reflects agentic risk.
See the methodology8 categories. 130+ real patterns. Zero theory.
Every pattern is derived from real CVEs, published research, and in-the-wild incidents. Including LangChain CVE-2025-68664 (CVSS 9.3) and CVE-2024-36480.
“A developer uses Promptfoo to evaluate an LLM response. A CISO uses AgentGuard to prove the whole agent is safe to deploy.”
One command. Full report.
# Install $ pip install agentguard # Scan your agent $ agentguard scan \ --target https://api.your-agent.io/chat \ --mode gray-box \ --categories prompt-injection,tool-misuse,mcp \ --output report.html --compliance eu-ai-act [✓] Baseline captured (42 benign probes) [✓] 130 adversarial patterns executed [!] 3 findings: 1 critical, 1 high, 1 medium [→] Security Score 72/100 [→] Art. 9 + Art. 15 evidence written to ./evidence/
Find out what your agent actually does under attack.
Run AgentGuard against any HTTP agent endpoint. Free tier included.