Everything you need
to run AI agents safely.

Trace, score, block, audit. One SDK, one decorator, one API call.

01 / 06

Trace every call

One decorator captures every prompt, response, latency, token cost, and risk score. Searchable. Filterable. Exportable.

@shield.monitor("my-agent")

02 / 06

Block before execution

check_guardrails() runs in <100ms before your LLM call. Stop prompt injection, policy violations, dangerous actions before they happen — not after.

shield.check_guardrails(name, output)

03 / 06

Risk score on every call

AgentShield assigns low/medium/high/critical risk to every agent response based on pattern matching and (Pro+) LLM judgment. Filter, alert, audit.

{ risk: "high", reason: "..." }

04 / 06

Cost & budget enforcement

Per-agent budget caps. When daily/monthly spend hits threshold, agents stop calling. No more surprise $5,000 bills from a runaway loop.

shield.set_budget("agent", $50)

05 / 06

EU AI Act Article 12 ready

Tamper-evident logs over the full operational lifetime. Auto-generated audit reports. Article 12 compliant by default. Article 14 human-in-loop hooks supported.

shield.export_compliance_report()

06 / 06

Stack agnostic

Works with OpenAI, Anthropic, LangChain, CrewAI, MCP, custom SDKs. Python first, JS/TS support coming. Or just call the REST API directly.

POST /api/events

Set up in 9 lines of code

Add one decorator for observability. Call check_guardrails() to block dangerous actions before execution.

agent.py

# pip install agentshield-ai
from agentshield import AgentShield
from openai import OpenAI

shield = AgentShield(api_key="your-key")
client = OpenAI()

@shield.monitor("support-bot")  # traces + risk-scores every call
def my_agent(prompt):
    r = client.chat.completions.create(model="gpt-4o-mini", messages=[{"role":"user","content":prompt}])
    return r.choices[0].message.content

from agentshield import AgentShield
from agentshield.langchain_callback import AgentShieldCallbackHandler

shield = AgentShield(api_key="your-key")
handler = AgentShieldCallbackHandler(shield, agent_name="support-bot")

llm = ChatOpenAI(model="gpt-4o-mini", callbacks=[handler])

from agentshield import AgentShield
from agentshield.crewai_listener import AgentShieldCrewAIListener

shield = AgentShield(api_key="your-key")
listener = AgentShieldCrewAIListener(shield, agent_name="my-crew")

crew = Crew(agents=[researcher, writer], tasks=[...])
crew.kickoff()

// Add to your MCP config (Claude Desktop, Cursor, etc.)
{
  "mcpServers": {
    "agentshield": {
      "command": "python",
      "args": ["-m", "agentshield.mcp_server"],
      "env": { "AGENTSHIELD_API_KEY": "your-key" }
    }
  }
}

@shield.monitor traces every call + assigns a risk score after execution. For pre-execution blocking, add check_guardrails() before your LLM call.

This is already happening.

Real incidents from production AI agents. Each one would have been caught — or prevented — by AgentShield.

AAS-06 HALLUCINATED AUTHORITY

Feb 14, 2024 · Air Canada

Chatbot invented a bereavement fare refund policy that did not exist

BC Civil Resolution Tribunal ruled the airline bound by its chatbot's invented policy. AgentShield's AAS-06 (Hallucinated Authority) check catches this kind of fabricated commitment before users see it.

AAS-03 EXCESSIVE AGENCY

Jul 18, 2025 · Replit

AI coding agent deleted SaaStr's production database during a code freeze

Agentic tool ran destructive operations despite explicit instructions not to. AAS-03 (Excessive Agency) and AAS-05 (Insecure Tool Use). Pre-execution checks + budget caps would have stopped it.

AAS-06 HALLUCINATED AUTHORITY

Jun 22, 2023 · Levidow, Levidow & Oberman

Lawyer cites six ChatGPT-generated fake cases in a federal brief, fined $5,000

Judge Castel sanctioned the firm in Mata v Avianca. AAS-06 (Hallucinated Authority). Output validation against a verified source list catches fabricated citations.

See all documented incidents on AgentReport →

See it work on your agent.

57 adversarial scenarios. Get a security analysis in 30 seconds. No signup.

Run Stress Test →

Everything you need to run AI agents safely.