Yes, a fake or lookalike online shop can mislead an AI agent. Agents read pages much as people do, so a convincing design, fake reviews or an unbeatable price can steer them. Worse, a page can carry hidden instructions that try to change what the agent does at checkout. Vendors defend with approval screens, single-use cards and separate action-checking classifiers, but none claims perfect protection. This guide is based on the official docs and launch coverage as of November 2026.

How fake shops target agents

Scam shops are an old problem, but agents change the economics. A person might notice a misspelled domain, while an agent hunting for the lowest price may rank the scam site first. Consumers reported $15.9 billion in fraud losses to the FTC in 2025, according to its testimony to Congress, and shopping agents add a new kind of buyer for scam sites to target.

Three traps stand out:

  • Lookalike domains and storefronts. Copied branding, a slightly altered domain and fake reviews. The agent may not weigh a new or strange domain as heavily as a careful shopper would.
  • Too-good deals. Prices far below market are a classic lure. An agent told to "find the cheapest" is exactly the buyer these sites want.
  • Checkout traps. Hidden text or page elements aimed at the agent, such as instructions to add items, switch the delivery address, or pay by bank transfer or gift card. This is a form of indirect prompt injection.

How vendors defend

Defence What it does Example
Per-purchase approval You see merchant and amount before paying Duet money actions always ask; Muse approves every purchase; Google says "We'll always ask for your permission first"
Separate action checker Reviews actions outside the main model Duet's classifier, where "Instructions hidden in a page cannot affect the outcome"
Single-use or scoped cards Caps the loss to one merchant and amount Link one-time cards (Duet, Muse)
Hard money limits Blocks transfers that scams push for Duet bans moving money between accounts; dots hands transfers back to you
Takeover for payments You enter payment details yourself Grok Bot

As of November 2026. Sources: Anthropic's Duet safety article, Google's agentic checkout post, vendor help pages.

These are meaningful defences. Anthropic reports a 1.4% attack success rate against Claude Opus 4.5 with safeguards in its prompt injection research, while saying "prompt injection is far from a solved problem". OpenAI's system card for GPT-6 Astra, the model dots launched on, reports an 8.5% attack success rate on the Gray Swan prompt injection benchmark, compared with 27.0% for GPT-5.6 Sol. Low is not zero.

Prompt injection success on Claude Opus 4.5 with safeguards (Anthropic)
1.4%
Gray Swan attack success, GPT-6 Astra (OpenAI system card)
8.5%
US fraud losses reported to the FTC in 2025
$15.9B

What agent verification does and doesn't solve

Payment networks are building ways for merchants to verify agents. Visa's Trusted Agent Protocol uses cryptographic agent signatures, and the IETF Web Bot Auth draft verifies the agent operator. These help genuine shops trust genuine agents. They do not tell your agent that a shop is genuine. Verification currently runs in the merchant's direction, not the buyer's.

Single-use cards, covered in agent payment security explained, are the main buyer-side protection: they stop repeat charges and overcharges. They will still pay a fake shop once if you approve it.

What to check on every approval

  1. Read the merchant nameDoes it match the store you expected, not a lookalike or unknown reseller?
  2. Check the amount and currencyWatch for added items, extra shipping or a different currency.
  3. Check the delivery addressA changed address is a red flag for a checkout trap.
  4. Question the payment methodBank transfers, crypto and gift cards are scam signals.
  5. Ask where the deal came fromIf the price is far below other stores, look yourself before approving.

Tell your agent to shop only at stores you name, where it supports that. An agent restricted to a few known retailers has far less exposure than one told to find the cheapest price anywhere. For how hidden instructions combine with access and outbound actions, see the lethal trifecta.

Warning signs your agent may not weigh

Agents are good at comparing prices and poor at the gut checks people make. Before approving a purchase from an unfamiliar store, look for: a domain registered very recently or one letter off a famous brand, no physical address or a mismatched country, reviews that are all five stars and posted in the same week, and payment options limited to bank transfer, crypto or gift cards. Look up the store name with the word "scam" or check it on a review site you trust. Thirty seconds of checking is enough for most orders.

If your agent bought from a fake shop

  1. Call your card issuer right away, explain the purchase was from a fraudulent site and ask for a replacement card.
  2. Report the site to the FTC at reportfraud.ftc.gov.
  3. If you lost money, file a complaint with the FBI's IC3.
  4. Report the action to the agent vendor through its help channel.
  5. Follow the 24-hour unauthorized purchase checklist for disputes and evidence.

Scammers also imitate the agents themselves; see fake AI agent apps. For a comparison of shopping agents and their checkout rules, see AI shopping agents compared, and for background on hidden instructions, prompt injection explained.

Bottom line

Fake shops can fool AI agents, and hidden page instructions make it easier. Vendor defences, including approval screens, action-checking classifiers and single-use cards, cut the risk without removing it. The approval screen is your last check: confirm the merchant, amount, address and payment method, restrict shopping to stores you trust, and report any scam to your card issuer, the FTC and IC3.