Gardien

Security

Agent security starts before the action runs.

One practical control point: should this agent, owned by this operator, take this specific action right now?

The model

Four layers, one enforceable answer.

01

Agent identity, not human identity

A passport describes the agent — owner, runtime, purpose, expiry, status. It never pretends an AI agent is a person.

02

Scoped action checks

The verifier evaluates the exact action requested: checkout submit, booking change, account update, API write.

03

Approval before blast radius

Low-risk actions pass automatically. High-risk actions wait for a human before money, data, or production changes.

04

Revocation-first trust

Trust must be easy to stop. Revoked agents stop receiving trusted decisions before the next sensitive action.

Verifier inputs

A useful checkpoint checks context, not vibes.

  • Passport ID and current status
  • Agent owner and runtime context
  • Requested action and protected surface
  • Origin, timestamp, nonce, replay risk
  • Policy match and approval requirement
  • Expiry, pause, and revocation state

Decision outcomes

Easy for a system to enforce.

allowedThe agent is active and the request fits policy, scope, freshness, and origin checks.
needs_approvalThe agent may be valid, but the action requires a human before execution.
deniedThe request doesn't meet policy, scope, signature, origin, or risk requirements.
expiredThe trusted window closed and a fresh decision is required.
revokedTrust was stopped; the protected action should not continue.

Honest boundaries

What Gardien is not.

Not a fraud guarantee

Gardien decides whether a known agent may take a scoped action — it doesn't promise the action is harmless.

Not a legal identity system

A passport identifies an agent for workflow trust. It's not KYC or government ID.

Not a replacement for auth

Keep your existing auth, fraud checks, rate limits, and human review.

Not a reason to log everything

Verifier logs support audit without storing raw records, passwords, or keys.

Security reports go to security@gardien.io with impact, affected route, reproduction steps, and timestamps. Never email live secrets, passwords, or customer records.

Next step

Map the first protected action before you trust agent traffic.

Protect a business actionRead verifier docs