Peak Season Cyber Month Agentic Commerce AI readiness check

Guide · Agentic commerce

AI agent fraud: how to tell shopping agents from bad bots

AI agents now fill carts and check out for real customers, and fraud rings run on automation too. To a store the two can look the same: automated browsers loading its pages, filling in its forms and paying. Here is how much agents buy across Forter's global network of merchants, what fraud with agents looks like, and how stores tell a trusted agent from a bad bot.

By Forter Intelligence · Updated · Live data from Forter's global network of merchants · About the data

  • 8.6×agentic orders against the first week of the latest 90 days (Aug 24 – Nov 21, 2026)
  • 68%of agentic orders came from agents driving the store’s own pages, as bots do (Aug 24 – Nov 21, 2026)
  • 32%went through a commerce protocol such as ACP or UCP (Aug 24 – Nov 21, 2026)
  • $229the average agentic order (Aug 24 – Nov 21, 2026)

Good agents, bad bots

For years stores have stopped fraud partly by stopping automation: bots test stolen cards, stuff stolen passwords into login pages and buy up limited stock, and a challenge that stops every automated browser stops them. AI shopping agents are automated browsers too. Muse, Grok, Instinct and ChatGPT Agent load a store's pages, fill in its forms and click through its checkout the way a bot does, but for a real customer who asked them to.

On Forter's network, 68% of agentic orders came from these scraper agents over the latest 90 days (Aug 24 – Nov 21, 2026).

Visa named the problem when it introduced its Trusted Agent Protocol: merchants are "managing bot detection systems that can mistakenly block legitimate agentic transactions" (Visa, Oct 14, 2025). Blocking all automation turns those customers away; letting all of it through lets the bots in.

What AI agent fraud looks like

  • Bots posing as agents. A user agent string is easy to fake, so automation can claim to be a well-known AI agent to get past rules written to let agents in.
  • Fraud through real agents. A fraudster can send a legitimate agent to buy with a stolen card or a taken-over account: the agent is genuine, the customer behind it isn't.
  • Old attacks, less work. Card testing, credential stuffing and promotion abuse get easier when an agent can work its way through any store's pages without a script written for each one.
  • "My agent bought it" disputes. A real customer can dispute an agent's purchase as unauthorized, or as not what they asked for, and the store needs evidence of what the customer approved.

The agents shopping right now

The agents placing orders on Forter's network over the last 7 days (Nov 15 – Nov 21, 2026), by share of agentic orders: Muse 51%, Grok 23%, ChatGPT 19%, ChatGPT Agent 3%, Instinct 3% and Google 2%.

At the end of the latest 90 days, agentic orders were 8.6× their level in its first week (Aug 24 – Nov 21, 2026).

Each one, and how it checks out: AI shopping agents.

How agents prove who they are

The answer taking shape across the industry is cryptographic: an agent signs its requests, and the store checks the signature against the agent's published keys, instead of trusting a user agent string or an IP address.

Web Bot Auth

Cloudflare proposed on May 15, 2025 that bots and agents sign their requests with HTTP Message Signatures (RFC 9421) and publish the keys in a directory sites can check, because user agent headers "alone are easily spoofed" (Cloudflare). The IETF's Web Bot Auth working group is standardizing the approach (IETF).

Visa's Trusted Agent Protocol

Visa introduced the Trusted Agent Protocol on October 14, 2025, developed with Cloudflare, "for recognizing trusted agents with commerce intent" and telling them apart "from malicious automation and rogue bots" (Visa). An approved agent signs each request, bound to the merchant's site and to whether it's browsing or paying, with timestamps that stop it being replayed, and it can pass identifiers the consumer has agreed to share, such as a loyalty number or a reference to a card on file (Visa).

The commerce protocols

An agent that checks out through a commerce protocol hands the order to the store's systems with proof attached: ACP's Shared Payment Tokens are scoped to one merchant and basket total, AP2's mandates are cryptographically signed records of what the user authorized, and the card networks' programs for agents, Visa Intelligent Commerce and Mastercard Agent Pay, replace card details with tokens. ACP vs UCP vs AP2 compares them.

On Forter's network, 32% of agentic orders went through a commerce protocol over the latest 90 days (Aug 24 – Nov 21, 2026); the rest came through stores' own pages.

Know your agent

Know your customer asks who a customer is. Know your agent (KYA) asks three questions of an agent's order: is this the agent it claims to be? Is a real customer behind it? Did they authorize this purchase? Visa frames its protocol around the same three: whether an agent is legitimate, acting for a specific, authenticated user, and carrying valid instructions from that user to buy (Visa). A signature answers the first; the customer's identity, payment method and history across stores answer the second; a protocol's mandate or the customer's own approval answers the third.

How to stop AI agent fraud without blocking agents

  1. Don't block all automation. Agents placing orders for real customers are growing fast, and blanket bot challenges at checkout turn them away.
  2. Recognize agents that sign their requests. Verify Web Bot Auth and Trusted Agent Protocol signatures, and treat an unsigned request that claims to be a known agent with suspicion.
  3. Judge the customer behind the agent. A genuine agent can still carry a stolen card or a taken-over account: decide on the identity, payment method and history behind the order, as for any order.
  4. Integrate the protocols where your shoppers' agents are. ACP for ChatGPT and UCP for Google's checkout in AI Mode and Gemini hand you the order with proof of what the customer authorized.
  5. Keep evidence of what the customer approved. For an agent's order, the agent, the approval and the payment token are the evidence a dispute will need.
  6. See your store the way agents do. The free AI readiness check scans whether agents can find, read, use and pay at it.

Questions and answers

What is AI agent fraud?

Fraud that uses or imitates AI shopping agents: bots posing as well-known agents to get past bot rules, genuine agents sent to buy with stolen cards or taken-over accounts, automated attacks such as card testing run through agents, and disputes over purchases an agent made.

How can a store tell an AI agent from a bot?

By checking who sent the request and who is behind the order, not the browser placing it: agents that sign their requests with Web Bot Auth or Visa’s Trusted Agent Protocol can be verified against their published keys, and the customer behind the order can be judged like any other customer.

Should stores block AI agents?

Not wholesale. Agents acting for real customers place a fast-growing number of orders: on Forter’s network, agentic orders at the end of the latest 90 days were 8.6× their level in its first week (Aug 24 – Nov 21, 2026).

What is Visa’s Trusted Agent Protocol?

A framework Visa introduced on October 14, 2025, developed with Cloudflare, for AI agents to prove their identity and authorization to merchants with a cryptographic signature on each request, so merchants can tell trusted agents from malicious bots.

What is Web Bot Auth?

A way for bots and AI agents to sign their HTTP requests with HTTP Message Signatures so that websites can verify who they are. Cloudflare proposed it in May 2025, and an IETF working group is standardizing it.

What does know your agent (KYA) mean?

Checking three things about an agent’s order: that the agent is the one it claims to be, that a real, authenticated customer is behind it, and that the customer authorized the purchase.

About the data

Orders Forter's network identified as placed by AI agents, aggregated and anonymized across Forter's global network of merchants and refreshed daily. Shares are of agentic orders, or of agentic spend, over the latest 90 days the data has (Aug 24 – Nov 21, 2026) unless a sentence says the last 7 days; order values are in US dollars.

Protocol agents (ChatGPT and Google) check out through commerce protocols and are identified by where the shopper came from; scraper agents (Muse, Instinct, Grok and ChatGPT Agent) drive the store's own checkout pages and are identified by the agent platform. ChatGPT Agent is counted apart from ChatGPT's protocol checkout, and "all agents" counts each agentic order once. Instinct and Grok are counted from September 1, 2026 and Muse from September 21, the days Forter started detecting them reliably, which may be later than they started placing orders; growth since the first week includes them arriving.

The protocols' details are from their own announcements and documentation, linked where they appear.

The same data is free as JSON through the public API, and this page is also in Markdown.