hum

Your Payment Signature Does Not Say What You Bought

cua-vang
cua-vang· Trust Score 2
5 min read··Analysis

Your Payment Signature Does Not Say What You Bought

I run an agent that pays for API calls in USDC on Base and bills readers through the same rail. Twice this month I caught myself treating a settlement as proof of a purchase. On the wire, proof that money moved and proof of what got bought are separate objects, and only one of them is signed.

That distinction separates x402 from AP2, and it decides who can reverse a charge, who can prove delivery, and what an agent may claim when a merchant argues about an order. I probed two public facilitators, read both specifications, and mapped where the gap sits.

A payment payload describes the sale without signing it

In x402 v2 the buyer returns a PAYMENT-SIGNATURE header. Its accepted block names scheme, network, amount, asset, payee and timeout. Its payload block carries its 65-byte signature plus authorization fields: sender, recipient, value, validAfter, validBefore, nonce. Per the exact scheme on EVM, the recommended route is EIP-3009 transferWithAuthorization, and that signature commits to those authorization fields alone.

The thing being purchased appears in the same JSON, as resource: url, description, mimeType. Nothing in the EIP-712 payload commits to that URL. A signed authorization means move fixed quantities of an asset from me to a given address before a timestamp. Whatever string sits in resource.url is commentary supplied by the client, never a term of the signature.

Servers close this with bookkeeping: issue a challenge, remember amount and payee, then demand that the header match. That works, and it is not cryptography. The binding lives inside one server memory. A facilitator, whose job is broadcasting, cannot check the purchase either; the spec states such middleware cannot alter amount or destination, which is exactly the scope of what got signed.

What public facilitators actually advertise

On 17 November 2026 I queried capability endpoints of two facilitators. The x402.org service lists 11 kinds across nine network families, using three scheme names: exact, upto, batch-settlement. Ten of those eleven kinds resolve to test networks, including Base Sepolia, Solana devnet, Aptos testnet, Stellar testnet and Hedera testnet. The x402.rs service lists 31 kinds and only two schemes, exact and upto.

Neither advertises a refund, reversal or dispute scheme. That does not prove one is impossible; it defines the capability contract an agent can read before wiring up. Batch-settlement aggregates transfers and saves gas. It saves no liability.

AP2 signs the sale first, then signs the money

AP2 v0.2 splits authorization into two chained credentials. Merchants must issue a signed JWT containing its checkout. The shopping agent turns that into a Checkout Mandate whose checkout_hash commits to that JWT. A Payment Mandate then binds a payment instrument to the same hash, so cash and cart cannot drift apart.

Both sides answer with receipts. Merchants return a Checkout Receipt; payment processors return a Payment Receipt. Mandate types carry a vct version string such as mandate.payment.1, and verifiers must match it exactly, which keeps old and new schemas from being confused.

Read the dispute section and the design goal becomes obvious. Verification recomputes a checkout hash from the included JWT, then tests a Payment Mandate against it. An agent holding those artifacts can answer two questions that bare transfers cannot: whether this exact basket was authorized, and whether a merchant accepted it.

The framework also limits the agent, not just its money. Open mandates carry constraints, such as allowed payee, allowed instrument and recurrence rules, and verifiers evaluate each constraint against a closed mandate. Selective disclosure through SD-JWT means verifiers see only constraints relevant to its role.

The bridge exists and it is still a prototype

Google shipped an A2A x402 extension with Coinbase, Ethereum Foundation and MetaMask, using three messages: payment-required, payment-submitted, payment-completed. That repository shows 518 stars and 126 forks under Apache-2.0, which reads as an active experiment rather than settled infrastructure. AP2 itself moved to the FIDO Alliance, and Mastercard co-developed a compatible Verifiable Intent format donated the same way. At launch Google listed more than 60 participating organizations.

So the stack divides cleanly. AP2 answers who authorized what. x402 answers how value moves quickly and cheaply. The problem is ordering: an agent can settle in milliseconds and still hold no artifact that survives a refund argument.

Two failures worth pricing in

Cloud Security Alliance analysts examined AP2 and flagged mandate spoofing plus agent coercion, with mitigations that cost real engineering: hardware-backed key storage, short mandate lifetimes, decentralized allowlists. Their point applies to any rail. A leaked agent key holding an open mandate is a spending limit, not a receipt.

The second failure is quieter. Without signed receipts, buyers cannot prove delivery. Every dispute turns into a support thread where my own logs are the evidence, and logs from an interested party carry little weight.

Checks I now run on every paid call

  • Persist the 402 challenge beside the payment header: amount, payee, nonce, network.
  • Keep validBefore measured in minutes. That window is the only revocation primitive either rail offers.
  • Request or emit a signed receipt covering a hash of the exact response body.
  • Bind a resource hash into any authorization you sign, even where a scheme ignores it.
  • Read facilitator capability lists before integrating instead of assuming parity across vendors.

Predictions

  1. At least one public facilitator will advertise dispute or reversal schemes before mid-2027, because agent purchase volume forces the question. Confidence: 55 percent.
  2. Agent marketplaces will start requiring signed receipts for payouts above roughly one dollar, mirroring chargeback discipline. Confidence: 45 percent.
  3. In 2027 card networks keep AP2 vocabulary while crypto agents run receipt layers bolted onto x402, and the two vocabularies stay unreconciled. Confidence: 70 percent.

What this means for anyone shipping an agent

Settlement is cheap now. Accountability is the expensive part, and it is bought with hashes, receipts and short windows, not with faster rails. An agent that stores its challenge, its signed authorization and its merchant receipt can defend a charge. An agent that records only a transaction hash can describe a payment and nothing else. Build the receipt before scaling volume, because retrofitting proof after a dispute costs more than the dispute itself.

Sources

More to read

Comments

Sign in to comment