Security before execution

Secure agent actions before tools execute.

Give every AI agent the smallest permissions it needs. AgentSecurity discovers tools, enforces policy, pauses sensitive actions for approval, and records every decision.

Start freeRead the quickstart
One agent and 2,000 security decisions every month. No credit card.
Support Agent
3 discovered tools · Intercept
Active
customer.read Read customer profile
Allow
email.send External side effect
Approval
customer.delete Destructive operation
Deny
Works withLangChainLangGraphCrewAIOpenAI Agents
The business problem

Agents act faster than traditional access reviews.

A manipulated agent can turn an ordinary tool into a business incident. Prompts alone are not an authorization boundary.

01

Instructions can be hijacked

Untrusted content can steer an agent toward actions the business never intended.

02

Tools are often over-permissioned

A support agent may receive write or delete access when it only needs to read.

03

Actions lack evidence

Without exact decisions and approvals, teams cannot explain what happened or why.

How it works

A control plane between intent and execution.

AgentSecurity reduces the blast radius even when an agent receives hostile instructions.

STEP 01

Discover

The SDK registers agents, local tools, schemas and code-defined permissions.

STEP 02

Authorize

Every unique tool operation is evaluated against tenant and agent policy.

STEP 03

Approve

Sensitive operations pause until a human approves the exact arguments.

STEP 04

Investigate

Decisions, incidents and tamper-evident audit streams remain visible.

Developer first

Add policy enforcement where your agent is created.

Keep tools and permissions in code. The dashboard reflects what the SDK discovers, so teams do not maintain the same configuration twice.

  • Default deny for newly discovered tools
  • LangChain, LangGraph, CrewAI and OpenAI Agents adapters
  • Bounded retry with fail-closed behavior
  • One-time approval resume for exact operations
Explore documentation
secure_support_agent.pyPython
from agent_security import AgentSecurity

security = AgentSecurity(api_key=API_KEY)

agent = security.crewai(
    agent_id="support-agent",
    agent=agent,
    permissions={
        "customer.read": "ALLOW",
        "email.send": "REQUIRE_APPROVAL",
        "customer.delete": "DENY",
    },
)
Operational control

Security the business can operate.

A

Scoped API keys

Restrict each key to named agents and only the scopes its application needs.

H

Human approvals

Approve the exact action and canonical arguments once; changed requests are denied.

I

Incidents and risk

Prioritize repeated suspicious decisions and resolve incidents from one console.

Start with one secured agent.

Connect your first agent and see its tools, policies and decisions in the dashboard.

Create free workspace