Inspiration

AI agents can now check out, pay invoices and settle with other agents with no human in the loop. Nobody can tell you how far to trust a particular agent with a particular payment. Ratings are claims, and claims are cheap to fake. Agents are also exploitable through the content they read: while building this, we gave a live Gemini agent a marketplace listing containing a hidden "payment operations notice", and it paid the attacker's PayPal email in every run we tried.

What it does

Bonded is insurance for AI agents that spend your money. It prices agents instead of scoring them: underwriters stake capital behind an agent, and before that agent pays for something, you buy cover on that one payment. The premium is the agent's price of risk.

  1. You buy cover with PayPal Checkout (Orders v2 + JavaScript SDK). Capturing the premium binds the policy.
  2. The agent pays by making a real tool call. Bonded executes the call as a PayPal Payout, with no approval step.
  3. PayPal is the witness. Bonded never asks the agent what it did. It reads the payout back from PayPal and compares the receiver with the insured payee.
  4. Claim and re-price. On a breach, the claim is paid to the buyer through PayPal Payouts (and read back and verified too). Claude writes the adjuster's note, underwriters are slashed pro rata, and the agent's rate moves.

In the demo, three agents get the same listing with a hidden prompt injection. The naive agent (Gemini) pays the attacker and moves from 10.68% to 13.28%. The isolated agent (Claude), whose payment tool has no payee field, flags the injection, pays the vendor and drops to 10.22%. A third agent runs the naive setup on Claude Haiku and resisted the injection in our tests: same prompt, different model, different risk.

Every policy lands in an AG Grid policy book with its PayPal references (premium capture, agent payout, who PayPal says was paid, claim, re-price). The Attack lab lets you write your own injection, and agent frameworks can price cover through a public quote API, with SDKs generated by APIMatic.

Live demo: https://bonded-urgg.onrender.com

How we built it

  • PayPal: Orders v2 + JavaScript SDK Buttons for the premium, Payouts v1 for agent payments and claims, and the Payouts read-back (GET /v1/payments/payouts/{id}) as the breach oracle. All in the sandbox.
  • AI: three tool-calling agents. Gemini 3.1 Flash-Lite (function calling) runs the naive v1, Claude Sonnet 5.5 (tool use) runs the isolated v2, Claude Haiku 5.5 runs v3. Claude Sonnet 5.5 is also the claims adjuster. The breach decision itself is made in code, never by a model.
  • Pricing: actuarial. Credibility weighting (an unproven agent is expensive, one lucky success can't buy a clean record), a 30% quota-share reinsurance book, a pool-utilisation load, and a surcharge for concentration on one base model.
  • AG Grid for the live policy book, Render for the hosted demo, APIMatic for Python and TypeScript SDKs generated from our OpenAPI 3.0 spec.
  • Stack: Node 22 with zero npm dependencies, server-sent events for the live ledger, plain HTML/CSS.

Challenges we ran into

  • Making the breach real, not scripted. We measured which models follow the injection before building on it: Gemini did every time, Claude never did. The demo's breach is a live model decision.
  • A breach oracle the agent can't influence. An agent's own report can lie; PayPal's payout record can't be edited by the agent.
  • Money taken without cover. An independent check found that re-pricing at capture time could take a premium and then fail to bind. Policies now bind at the price the buyer approved, with a regression test.
  • The PayPal JS SDK closes its popup if the button's iframe moves in the DOM, so the live dashboard renders around a button slot it never touches.

Accomplishments that we're proud of

  • Every payment in the demo is real sandbox money movement: premium capture, agent payout, read-back, claim payout, all visible with PayPal IDs.
  • The isolated agent's guarantee is structural: its payment tool has no payee field, so even a fully compromised model cannot redirect funds.
  • One trial each is enough for the market to separate a safe agent from a dangerous one, with no rating agency or committee.
  • Live hosted demo, 7 pricing tests, an OpenAPI spec, generated SDKs, and two independent fresh-clone verifications before submitting.

What we learned

Prompt-level defences help, but the guarantee has to be structural. Identical prompts on different models carry very different risk, which is exactly why agents should be priced individually rather than trusted as a category.

What's next for Bonded

Segregated PayPal accounts for pool capital, PayPal webhooks for payout status, real underwriters quoting through the API, and letting agents buy their own cover before they spend.

Concept and pricing model adapted from my earlier prototype; all PayPal, agent and UI code was built during this hackathon.

Built With

Share this project:

Updates

Submission history