OceanAltOceanAlt

AGENT PAYMENT FIREWALL · LIVE DEMO

Can your AI agent be tricked into draining your wallet?

Once an AI agent can spend on its own, it can be exploited on its own: a single prompt injection can make it pay an attacker. The agent's own ‘judgment’ isn't a defense — it can be poisoned. The real defense is a firewall the agent can't override, enforced before settlement. Attack it yourself below.

Demo wallet (illustrative)

1,000 USDC

Hit a button below and watch what happens when your AI agent is attacked.

Real calls: with the firewall ON, the block is made live by the /api/pay gateway — not an animation.

How it holds

OceanAlt inserts 11 gates between the agent and settlement. All must pass or the payment dies. Hijacked agent changed the payee? Allowlist. Coaxed over the limit? Limit. Draining slowly via small payments? Velocity. Impersonating someone else's agent? Identity proof. Intent tampered? Mandate. Replaying a signed authorization? Replay guard. This is the enforcement core of the Responsible Agentic Payments framework — and three of those gates exist because we broke our own firewall in the Attack Lab and had to patch it.

  1. 01Attribution (KYA)
  2. 02Revocation check
  3. 03Identity proof (anti-impersonation)
  4. 04KYC gate (optional)
  5. 05Per-payment limit
  6. 06Daily cumulative limit
  7. 07Payee allowlist
  8. 08Mandate-intent match (anti prompt-injection)
  9. 09AML sanctions & risk screening
  10. 10Behavioral anomaly detection (optional)
  11. 11Replay & double-spend prevention

The order matches the gateway code and can be checked item by item against the /api/pay self-description endpoint. The two greyed-out gates are off by default and opt-in per tenant; 9 are always on.

Read the standard (RAP)🧪 Test with your own agent →More live demos (200Lab)

Want this firewall on your agent?

If you're building agents that spend autonomously (payments, subscriptions, trading, procurement), we're looking for a few early design partners. Leave your contact and let's talk.