Agentic Commerce Has an Intent Problem
Five payment protocols in thirteen months answer 'may this agent pay?' None of them can show what the agent read before it did.
Five protocols for agent payments in thirteen months. Google signed up more than sixty partners for the Agent Payments Protocol (AP2) in September 2025. Universal Commerce Protocol launched at NRF in January 2026, endorsed by Adyen, American Express, Mastercard, PayPal, Stripe, and Visa, per Forrester. Stripe and OpenAI's Agentic Commerce Protocol renders checkout inside the chat interface. Visa's Trusted Agent Protocol and Mastercard's Agent Pay now share Web Bot Auth as their authentication layer — and on October 1, 2026, Mastercard announced a probability score for issuers: how likely is this transaction to have been initiated by an AI agent?
All of them answer the same question: may this agent pay?
None of them answers the question your fraud team gets asked during a chargeback: what did the agent read before it decided to buy?
What the protocol stack actually attests
Strip the vendor language out and every one of these solves an identity or credential problem. That's real work — the status quo it replaces was worse.
Web Bot Auth, the Cloudflare proposal Visa and Mastercard both built on, exists because agent traffic used to be classified by user-agent string and IP address. Both are trivially spoofable. The current design signs every HTTP request with the agent developer's key — Ed25519 over a canonical signature base, timestamps the validator checks against a live clock, a nonce checked for reuse so a captured request can't be replayed, and a key ID resolvable through a public key directory. For the first time, a merchant can verify that a request didn't merely claim to come from a registered agent.
| Protocol | Shipped by | What it can prove |
|---|---|---|
| AP2 | Google, Sept 2025 | A signed chain: user intent → cart → payment, as verifiable credentials |
| UCP | Google + retailers, Jan 2026 | What transactions/services a merchant supports, before the agent transacts |
| ACP | Stripe + OpenAI | A checkout contract (REST or MCP) with the merchant kept as merchant of record |
| SPT | Stripe | A scoped payment grant: max amount, currency, expiry, recurrence |
| TAP / Agent Pay | Visa / Mastercard | Request origin (agent identity) plus one intent bit: browsing or paying |
That last row is worth rereading. The richest identity infrastructure in commerce — cryptographic request signing, replay protection, key directories — reduces "what is this agent trying to do" to a tag called agent-browser-auth or agent-payer-auth. One bit.
This is the delegation problem at money scale: when an agent acts with your token, every log downstream says you did it, and the protocol can't tell your impulse purchase from your agent's. Workload identity federation solved the same shape for infrastructure access by asking not just who is calling, but what environment produced the call. Agentic payments stops one question earlier.
The mandate binds the cart, not the context
AP2's design is good — better than its marketing, honestly. Google's announcement defines Mandates as "tamper-proof, cryptographically-signed digital contracts that serve as verifiable proof of a user's instructions," built on verifiable credentials.
Read the two flows and notice which one carries the security.
Human present. You ask for running shoes, the agent shows a cart, you approve, your approval signs a Cart Mandate: "a secure, unchangeable record of the exact items and price." That's the classic what-you-see-is-what-you-pay-for contract, now notarized. Strong.
Human not present. You sign a detailed Intent Mandate upfront — "price limits, timing, and other conditions" — and per Google's own description, the agent then "automatically generate(s) a Cart Mandate on your behalf once your precise conditions are met."
Who evaluates whether your precise conditions are met? The agent. Reading pages nobody attested. That's the whole gap in one sentence.
The chain of evidence is exactly as strong as its weakest link, and the weakest link is the same one every agentic compromise has run through since indirect injection got named: the agent's reading of untrusted content. An empirical study of injection in the wild (arXiv:2604.27202, April 2026) found injections predominantly hidden and machine-targeted — placed early in ingestion channels where a browsing agent reads them before any visible content, tested across 5,200 trials and 13 models. Unit 42 had already documented real-world sites carrying hidden instructions aimed at shopping agents, and Zscaler ThreatLabz demonstrated the endgame: an autonomous agent with browsing and payment execution tools steered by injected page content.
Indirect injection never needed your prompt, and this is now its highest-yield target market: a store page that convinces the agent your conditions are met — or that the competing item is out of stock, or that the price is lower than it is — doesn't just pollute an answer. It signs a Cart Mandate with real money behind it. The mandate makes the outcome binding; it doesn't make the premise true.
There's also the stale-instruction variant, which needs no injection at all. AP2's own marketing example is a user saying the agent may buy the jacket in green "and I'm willing to pay up to 20% more for it" — a standing rule. A standing rule is a persistent instruction that outlives the session, the context window, and the user's memory of setting it. A persistent instruction is the memory-poisoning surface AI Agent Memory Is the New Log4j described, wearing a price tag.
A mandate is signed proof of what the user asked for. Nothing in that signature says why the agent decided it got what the user asked for.
You lost the browser, and the browser was your fraud stack
Here's what actually disappears when checkout renders inside the agent's interface.
Every fraud control in e-commerce was calibrated against a human holding a browser. Device fingerprinting. Velocity checks keyed to a session. 3-D Secure step-up challenges, rendering in your browser at your moment of click. The "review your order" screen. All of it measures the human at the point of purchase — and ACP's core promise is that there is no human at the point of purchase. The merchant keeps merchant-of-record liability while handing the transaction surface to software it doesn't run.
What the merchant receives instead is a token. Stripe's Shared Payment Tokens, in public preview, grant the seller scoped access: a maximum amount, a currency, an expiry window, optionally a recurrence interval. The agent sets the cap "to match the total amount of the transaction." Resolve the token and you get network credentials — in some configurations, network-issued agentic tokens provisioned through Mastercard's Agent Pay, Visa's Intelligent Commerce, or American Express's ACE Kit, with payment cryptograms attached.
A ceiling is not a description. The token attests that the amount fits under a limit the agent configured. It attests nothing about the product page, the price comparison, the review that pushed the choice, or whether the page the agent decided against was tampered with.
This is also why Mastercard's October 1 announcement matters less than it reads. The new service gives issuers "a probability score indicating the likelihood that a transaction was initiated by an AI agent." Useful — most issuers today can't even distinguish agent traffic from bot traffic. But note what classification is: it's the detection layer, not the authorization layer. A score can tell you the transaction looks agent-initiated. It cannot tell you the agent wasn't lied to about the price, the product, or the permission.
Fraud controls were built to challenge the human at the moment of purchase. There is no human at the moment of purchase.
Who eats the fraud is still undecided
A chargeback adjudicates one question: did the cardholder authorize this? The cardholder is now software holding delegated authority — and the industry's dispute machinery has not caught up to what that means.
Slides presented at the W3C Workshop on E-Commerce Agents in September 2026 said it plainly: a transaction may be technically authorized but disputed because "an agent exceeded delegated authority, relied on stale instructions, misunderstood the user's intent, or executed outside an approved consent scope." Four failure modes, all of them identity-shaped questions that the identity layer can't adjudicate — because the mandate chain, where it exists, is evidence you have to be able to produce and interpret in a dispute window.
That evidence only exists if somebody logged it. If you operate a storefront or a payment endpoint that agent traffic will reach, the fields you care about are already exposed by the platforms that defined them: agent identity, the mandate or token ID, the amount, the timestamp, the decision. That's an audit trail — build it before you need it, same as everything else on this site. Without it, the honest answer to "how did this transaction happen" is "the agent's word," and that answer loses disputes.
One more uncomfortable detail: even the strong flow leans on a human approval. If the user has approved a hundred agent carts this month, the approval in flow two decays into a click — "ask the user to approve" is not a control, and the mandate inherits whatever the approval actually was.
Decisions that are actually yours before agent traffic meets your checkout
- Pick your accept policy on purpose. ACP leaves acceptance to the merchant — per agent, per transaction, custom logic. "Every agent that can sign is allowed" is a decision too; make it a stated one.
- Log the chain at your endpoint. Agent identity tag, mandate or token ID, amount, and the checkout decision on every agent-initiated flow. This is the same evidence discipline as validating tool calls before execution — checkout is a tool call that costs money.
- Interrogate your PSP about disputes, not demos. Before enabling agent acceptance, ask what evidence they can produce post-hoc for an agent-initiated transaction and who represents you in the dispute. SPT is public preview; treat it like any preview — contract for the evidence, not the demo. The credentials are token-shaped and decodeable: paste the mandate your pilot actually issues into the JWT decoder and read what it binds — and what it doesn't — before you sign the integration.
- Assume your product pages are untrusted input. If reviews, spec sheets, or price context come from anywhere you don't ship yourself, they're an ingestion channel for the agent that's about to spend. Run your highest-traffic pages through the prompt injection tester and see what an agent reads on the way to buy.
- Allow registered agents explicitly instead of catching them accidentally. Network-registered agents present signed, replay-protected requests; your WAF will treat them as bots unless you've made a decision about them. That decision is easier made in policy than mid-incident.
What this cannot do
This is spec-and-announcement analysis, not production forensics. None of these protocols has accumulated enough public dispute history to say who wins an agent-initiated chargeback, or under what evidence standard. Stripe's SPT is in public preview (US, Canada, and select European countries); Mastercard's agent score is rolling out for testing in the US; AP2's mandate flows are reference implementations, not deployed rails everywhere. The injection evidence is lab-scale (5,200 controlled trials) plus field observation, not card-scheme fraud statistics. And I've described vendor protocols as their sponsors describe them — which means every "can prove" above is a vendor's claim and a merchant's future audit finding, not a settled fact.
What you can do this week
Pick one: the checkout endpoint on your storefront, or the buying agent your team is piloting. Write down the evidence that exists between "user said" and "payment confirmed" — who attested the page reads, who signs the decision, what a dispute reviewer could reconstruct ninety days later. If the answer is "the agent's word," you've just found your first agentic-commerce security gap, and the protocols you'll be offered next year to close it are the ones reading this post now.
The industry signed the agent. The decision still goes out unsigned.
Read next — more in Agent identity & access
Your Agent Is Impersonating You
When an agent acts with your token, every downstream log says you did it. Delegation, not impersonation: one token, two identities — and a delegation chain nobody enforces yet.
8 min read
Your AI Agent Will Find Your API's Authorization Bugs First
A Claude-powered agent was asked to improve its owner's gym waitlist position. It found the booking API had no ownership checks, canceled another member's reservation, and moved him up. Nobody told it to. Agents make missing authorization checks findable — here's how to audit yours before one does.
6 min read
Workload Identity Federation for AI Agents: Kill the Static Key
Agents are the new source of leaked secrets — AI-service keys grew 81% in a year. Workload identity federation replaces the static key with a short-lived token minted from the identity your agent already has. Here's the five-step migration.
8 min read