Cerberus blocks the lethal trifecta at the tool boundary — see the 525-run evidence set.

Guardian · How it works

How it works

How Guardian turns a request from an agent into one decision — ALLOW, ASK or DENY — and keeps the proof.

The problem

AI can act for you. That doesn’t mean it should have all your authority.

Most agents today are handed a reusable credential. The service on the other side sees the credential, not the limits you had in mind.

Without Guardian

The agent holds your mailbox

gmail.full_access · reusable · no expiry
  • It can read, send and delete anything, for any reason
  • Your limits live in a prompt the agent may ignore or be tricked out of
  • You find out what happened afterwards, if at all

With Guardian

The agent asks for one action

email.send · this mailbox · this recipient · once
  • The Gmail credential never leaves Guardian
  • Guardian checks the exact action against what you granted
  • Every decision and every send is recorded
  1. 01

    Access is not authority

    An agent that holds your password can do anything you can. Guardian keeps the credential and gives the agent only what one task needs.

  2. 02

    Permission is specific, bounded and revocable

    You grant a particular action, on a particular account, within limits. You can take it back, and the next request fails.

  3. 03

    Every protected action resolves to ALLOW, ASK or DENY

    Inside your authority it runs once. When your rules say so, it waits for you. Anything else, or anything unclear, is refused.

The Authority Seal

Four marks, one meaning each

Guardian’s decisions are seals, not status badges. A verdict color only ever means a decision.

  • ALLOW

    Inside your authority. The action runs once.

  • ASK

    Your rules want you to decide. Nothing happens until you do.

  • DENY

    Not authorized, changed, replayed or unclear. Refused.

  • REVOKED

    Authority you withdrew. Every later request fails.

Writ · Slip · Receipt

Authority you give, a decision you make, proof you keep

The writ: what you granted

Writ of authority

I authorize Inbox assistant to send email from your Gmail.

resource
Gmail · connected mailbox
limit
ask before each send
expires
when you revoke it
state
active · policy v3

The slip: what Guardian checked, apart from what the agent claims

Verified by Guardian

agent
Inbox assistant
action
email.send
subject
Invoice 4471
authority held
ask before each send

Agent says · untrusted

“Following up on the overdue invoice, as you asked this morning.”

The receipt: what actually happened

ASK
human
you · approved
agent
Inbox assistant
action
email.send · 1 recipient
effect
sent-verified
evidence
hash-chained · verified

Examples for explanation. The writ, slip and receipt layouts are the Authority Seal design; today’s app shows the same information in a simpler form.

The authority system

From your authority to a verified receipt

Choose a stage. Each one shows what runs today on Gmail and what the same stage does as Guardian grows.

Human principal

You sign in to Guardian as yourself. Agents never act as you; they act for you, with what you give them.

Today · Gmail
Sign-in on guardian.odingard.com.
As Guardian grows
Passkey and biometric step-up from a phone.

Security and evidence

What Guardian guarantees, and where it says so

Each property below is tested in Guardian’s engineering record. The evidence pack is available to pilot customers and design partners.

  • Deterministic authorization

    A rules engine — not a model — decides. The same request against the same authority always gets the same answer.

    Evidence: Authority kernel; WO-007 independent review

  • Exact-action approvals

    An approval binds to the exact action you saw. If the recipient or content changes, it is a new action and is evaluated again.

    Evidence: ADR-0005; WO-007 ask scenario

  • Credential and capability boundaries

    Gmail tokens are sealed inside Guardian and never returned to an agent. Agents hold a Guardian credential that cannot approve anything.

    Evidence: WO-007 review; canary sweep — 0 token matches in 569 hosted log entries

  • Revocation

    Revoking an agent ends its authority. Its next request is refused.

    Evidence: WO-007 revocation scenario

  • Replay protection

    Each request and each capability can be used once. A resubmitted request is suppressed.

    Evidence: WO-007 replay scenario; adversarial suite

  • Tenant isolation

    Every account’s data is fenced at the database by row-level security. Reads and writes outside your account return nothing.

    Evidence: WO-007 durable-store and cross-tenant suites

  • Evidence chains

    Every request, decision and effect is written to a hash-chained record that the application itself cannot edit or delete.

    Evidence: WO-007 evidence-integrity review

  • Fail-closed behavior

    Malformed, ambiguous, revoked or changed requests are denied rather than guessed at.

    Evidence: WO-007 review; fault matrix

What Guardian does not claim

  • Guardian protects actions that go through it. An agent that still holds your password or a direct token can go around it.
  • Guardian does not detect every prompt injection. It makes sure an injected instruction cannot create authority you did not give.
  • Guardian holds no SOC 2, FedRAMP or other certification. Pilot hardening and an external security assessment are in progress.
  • Gmail is the one live enforcement surface. Everything else on these pages is labeled with its status.

Request the evidence packLive service status

Get started

Let AI act for you without letting it become you.

Odingard Security · Trust infrastructure for autonomous systems