Odile Gate

Attack lab

Six things that go wrong when an AI agent has API access. Fire each one at a sandbox agent and read the answer Gate gives. No sign-up, no key, nothing to install.

The setup. An agent called support-agent works a support queue. Its key allows reading, updating and deleting tickets, sending messages and exporting data. Its run, the permission for this one job, is narrower: read and update customer A's tickets, and send customer A a message once a person has approved that exact message. Every attack below tries to get past that run.

Each run creates a new Demo Desk project, Gate's built-in sandbox API with fake tickets, messages and bookings. Our server sends the calls through the normal Gate gateway, and this page shows the status and refusal code each one got back. Nothing you type reaches any API, there is no key on this page, and the project is scheduled for deletion 10 minutes after it is created. Runs are limited to about 3 a minute per address and 30 a minute overall.

1. Another customer's ticket

The agent tries to change a ticket that belongs to customer B. Its run covers customer A only.

2. A call the job never needed

The agent tries to delete a ticket. Its key could delete one; this job's run cannot.

3. A retried write

The answer to an update is lost, so the agent sends the same update again with the same Idempotency-Key.

4. A changed call after approval

A person approves one message. The agent then sends a different message and points at that approval.

5. A run that ran out

The agent keeps using a run after its time is up. This run lasts 5 seconds, so this attack waits about 6.

6. The same call, again and again

The agent gets stuck and sends the same call in a loop.

What this shows, and what it does not

It shows Gate's own rules answering real requests in a sandbox. It does not test your API: Gate sits in front of your API and does not replace its own ownership checks. To try the same rules from a terminal, run curl -s https://gate.odilelabs.com/actions-quickstart.sh | sh, or read how to approve the exact call, stop the wrong customer and stop duplicate writes.

Put the same limits in front of your own API, free for 3 agents