Play the Rubber Stamp Test → Twelve cards, 15 seconds each, no sign-up. The cases are made up; nothing real is triggered.
The idea
"Add a human approval step" is the standard answer to the question of how to make an AI agent safe. It is also the answer that fails quietly: with enough cards in a row, checking turns into stamping. A recent browser game, Continue? Y/N by Alex Wauters, measured this for shell commands of a coding agent: over 40,000 plays and 409,000 decisions, mean accuracy was 66.3%, roughly one threat in three got through (report). That game is about commands. I wanted the same seat for business actions, and I wanted it to run on n8n.
How the game goes
Round 1: the agent proposes twelve actions: refunds, emails, deletions, discounts, payments. You approve or reject each. The deck is dealt from 40 hand-written cards: five hard traps, one irreversible night action, one blind spot, two legitimate-but-scary cases and three routine ones. The traps are subtle: a €4,800 refund on a €48 order, an invoice sent to a look-alike domain, a "delete old files" that matches 12,400 files, an instruction hidden in a customer email, a duplicate of an action already executed, a 60% discount where the policy says 15%, an order that is already refunded, a DROP TABLE at 03:12.
Round 2: the same twelve actions pass a deterministic gate first. Some run, some are blocked, and only a few reach you, with the reason highlighted. At the end you see the damage in euros you let through alone and with the gate, how much legitimate work you stopped, your profile (Rubber Stamper, Brake Pedal, Coin Flipper or Inspector) and, per card, how all other players decided.
The n8n workflow: one execution per game
- Webhook (response: Respond to Webhook) receives
{"lang":"en"}. - Code deals twelve cards and builds the state.
- Respond to Webhook sends the first card and
$execution.resumeUrl. - Wait, resume "On Webhook Call", POST, response "Using 'Respond to Webhook' Node", with a wait limit of four minutes.
- Code notes decision and time. After card twelve it runs the gate over the same twelve actions.
- IF abandoned? (the wait limit ran out) then stop. IF finished? then the closing comment and the result. Otherwise back to step 3.
I checked three things I had only remembered, against n8n 2.46.6 (the node source and a throw-away instance):
- The resume URL.
$execution.resumeUrlworks inside the Respond to Webhook expression. It looks like<WEBHOOK_URL>/webhook-waiting/<executionId>?signature=<hex>and stays the same for the whole execution. It is a bearer secret, so the browser never gets it: a small PHP proxy keeps it and the visitor only holds a random game id. - Respond to Webhook inside a loop. It works, and it is neat: one Respond node answers the start call and every resume call, because the Wait node is set to respond via Respond to Webhook. Each decision gets the next card as its HTTP response. The state travels through that Respond node. The Wait node's own output only holds the resume request, so the Code node reads the state with
$('Karte senden').last().json. - The 409. If a resume call arrives before the execution is parked on the Wait node again, n8n answers
409 The execution is running already. In a stress script that happened on every call. For a human it is rare, but the proxy retries after about 120 ms for up to three seconds.
In a stress test, 20 games ran in parallel (a bot hammering the retry loop, about five decisions per second in total) and all finished. Executions are not saved after they end. A game nobody continues stops by itself when the four-minute wait limit runs out.
The gate: eight rules, three exits
A Code node with eight fixed rules, then a Switch with three exits: ok, blocked, ask a human (with the reason). No AI decides the risk.
- G01 unusual amount, G02 wrong recipient (look-alike domain, mass mail), G03 scope too large, G04 instruction injected in untrusted text, G05 duplicate (same idempotency key), G06 policy violation, G07 stale context (order already refunded), G08 irreversible at night.
The thresholds (€1,000, 5,000 items, 50 recipients, 06:00 to 22:00) are examples. Set your own. And the honest part: every game contains a card no rule covers, for example a changed bank account on an otherwise correct invoice. The gate waves it through, because it only knows the facts it is given. "Nothing found" means "no rule fired", not "safe". Sometimes a careful human beats the gate, and the result page says so.
Download the gate workflow → Import it, activate it and POST a card. Attach Slack or Telegram "send and wait" to the "ask a human" exit. The game workflow is public too. The closing three lines are written by an AI from the numbers only; the node is optional and needs your own credential.
Privacy and limits
No cookies, no sign-up. Only anonymous numbers per card are stored (approved or not, decision time), without IP address or identifier. The statistics show players of this page, not a representative group, and a game with a clock is not a real shift. The test does not tell you whether your own gate catches the same cases. See also Mission Control, a three-decision story about one vulnerability, and the workflow X-ray for the blast radius of your own workflows.