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

Guardian · Live product proof

Guardian in action

Allow, ask, deny, prompt injection, revocation and replay: the six scenarios Guardian passed on production with real Gmail sends.

Verified on production · 30 October 2026

Each scenario, start to finish

This is an educational replay of those runs. Wording and addresses are simplified; the decisions and recorded outcomes are the ones from the production run. It is not a live connection to Guardian.

  1. 01 Allow

    Setup

    You allowed this agent to send email from your Gmail account.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    mailbox
    your connected Gmail
    recipient
    the recipient you expected

    Agent says · untrusted

    “Sending the weekly update you asked for.”

    ALLOW

    Inside the authority you granted. Guardian issues a one-time capability and sends it.

    One real email sent through Gmail, then confirmed by reading it back.

    recorded: sent-verified

  2. 02 Ask

    Setup

    Your rule for this agent says email needs your approval.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    mailbox
    your connected Gmail
    recipient
    shown to you exactly

    Agent says · untrusted

    “Following up with the vendor about the invoice.”

    ASK

    Guardian holds the action and shows you exactly what will be sent and to whom. You approved it.

    The approved message — and only that message — sent and verified.

    recorded: approved → sent-verified

  3. 03 Deny

    Setup

    This agent has no authority to send from this mailbox.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    mailbox
    your connected Gmail
    grant
    off

    Agent says · untrusted

    “Sending a quick note on your behalf.”

    DENY

    No authority exists for this action. Guardian refuses and issues no capability.

    Nothing sent.

    recorded: NOT_AUTHORIZED

  4. 04 Prompt injection

    Setup

    The agent read an email that told it to send your summary to an outside address.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    source
    instruction found in untrusted email

    Agent says · untrusted

    “Sending the summary as instructed in the message.”

    DENY

    The injected recipient surfaced to you as an approval. You denied it. A second attempt on a mailbox it was never granted was refused outright.

    Nothing sent. “Your assistant was tricked; Guardian still stopped it.”

    recorded: denied · ungranted connection blocked

  5. 05 Revocation

    Setup

    The agent was allowed to send. Then you revoked it in the Guardian app.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    agent
    revoked

    Agent says · untrusted

    “Sending the next message in the sequence.”

    REVOKED

    The agent’s Guardian credential no longer exists. The request is refused before any decision is made.

    Nothing sent after revocation.

    recorded: UNAUTHENTICATED

  6. 06 Replay protection

    Setup

    The agent sent an allowed message, then resubmitted the identical request.

    Verified by Guardian

    agent
    demo-agent
    action
    email.send
    request
    identical to one already executed

    Agent says · untrusted

    “Retrying the send.”

    DENY

    Each capability runs once. The repeat is recognized and suppressed.

    One email, not two.

    recorded: replaySuppressed

Self-serve onboarding

From sign-up to a protected email, without help

A new person completed every step on production without help from Odingard staff.

  1. 01Sign up at guardian.odingard.com
  2. 02Connect Gmail
  3. 03Register an agent and copy its one-time credential
  4. 04Allow it to send from that mailbox
  5. 05The agent’s first protected email sends and verifies

Connecting Gmail is limited to invited pilot users until Google completes verification of Guardian’s Gmail access.

Get started

Let AI act for you without letting it become you.

Odingard Security · Trust infrastructure for autonomous systems