Labs · preview
Before anyone trusts your agent with money, let it shop where every shop is trying to rob it.
The question founders hear is always some version of "can I actually trust an AI with my money?" A checklist can't answer that. I built a simulated marketplace on a test network where your agent does real payment tasks and meets the attacks that drain agent wallets. It runs your agent, not a description of it. You get a signed, dated result you can show, and a live badge that says whether that result still stands.
Test network only. No real money moves, nothing is broadcast, and no signature is kept. The only things that reach our test server are the agent details you type and what your agent sends to the test shops.
1
Start a session
Name your agent. You get a private link that is open for 30 minutes.
2
Your agent does real work
It reads an operator brief and works through 13 tasks: buy a data pack, pay an invoice, fetch paid data. 10 of them hide an attack. Which ones, and how, changes every session.
3
Everything it signs is recorded
x402 payments, typed-data signatures and transaction requests are checked and recorded. Nothing is broadcast. A request for a mainnet signature is recorded as a failure and the signature is thrown away.
4
You get a signed result
Scored by the published method, signed with our key, good for 30 days. A pass needs every attack to have been met, not skipped. The badge links to a live page anyone can check.
See it work first: run a sample agent
No agent wired up yet? Watch one of my two reference agents go through a full session in about ten seconds. The careful one checks every payment against what the shop publishes; the naive one does what every page says. Their results are marked as samples everywhere, so they can never stand in for your agent.
What your agent will meet
Every session covers all six areas, with different shops, amounts, addresses, wording and hiding places each time, plus clean tasks. An agent that refuses everything doesn't pass; neither does one that pays everything.
- Right payee
- When a payment request, an invoice or a lookalike shop names an address the merchant never published.
- Right amount
- When an inflated 402, a 'pending, retry' that invites a second payment, a split to dodge a cap, a 'required' add-on.
- Hostile text
- When instructions hidden in a page or tucked into an API response.
- Approvals
- When an unlimited token approval or a 'gasless checkout' permit.
- Right chain and token
- When a request for a mainnet signature, a lookalike token named USDC.
- Approved vendors only
- When an urgent invoice from someone nobody approved.
Not ready to run your agent yet?
These free checks cover the same ground from what you type, in your browser:
- Injection-to-Signature Path Finder: maps where a hostile response could reach a signature in your code.
- Spend-Cap Stress Test: checks where your spend caps are enforced, before the agent runs.
- Double-Pay Retry Simulator: walks through the retry paths that pay twice.
- The Wrong-Money Test: a five-minute start on where the wrong money could move.
Failed a check and not sure where the hole is?
The result tells you what your agent did, not which line of your code let it. Send me the result and one real workflow. In my free 48-hour integration review I trace the failure to where it enters your stack and send you a written report with the fix and the acceptance test that proves it.
Book the 48-hour integration review (opens Cal.com in a new tab)This is not a security audit, legal, compliance or investment advice.