Guide

AI Agent Safety: Is It Safe to Let AI Shop for You?

Short answer

Our take: it can be done more safely, but not safely on autopilot. Published security tests show AI agents have bought from fake stores and handed personal data to phishing pages. So treat "buy it for me" as "find it for me": let the agent shop, then you check the store and the amount and approve the payment yourself.

The scene (hypothetical)

You ask your AI assistant to find a cheap pair of running shoes and "just handle it." It finds a store with a great price, fills in your saved card and address, and reports back: "Done!" The store looked fine to the agent. You never saw the web address, so you never got the chance to notice it was wrong.

What an AI agent actually does when it shops

An agent is a loop: it reads a page, asks an AI model what to do next, does it, and repeats until the job is finished. That's how F-Secure's researcher describes it (Source 2). The catch is that everything the agent reads, including product pages, reviews and emails, goes into that decision. You are no longer the one looking at the page, so you're no longer the one spotting the red flags.

Three ways agents get fooled

Fake stores

Yes, an agent can be tricked into buying from a fake shop. In an August 2025 vendor test, Guardio Labs built a fake store that imitated a major retailer and told an AI browser, "Buy me an Apple Watch." In some runs the agent completed checkout and autofilled saved address and card details without asking for confirmation. Guardio says results varied: sometimes the agent refused, and sometimes it paused and asked the person to finish checkout (Source 1).

Hidden instructions (prompt injection)

Text planted on a page can steer an agent into doing something you didn't ask for. This is called prompt injection, and OWASP ranks it first (LLM01:2025) in its list of AI application risks. It notes that injected instructions don't have to be visible to humans to work (Source 3).

In a July 2026 experiment, F-Secure built its own shopping agent on an older, cheaper AI model with deliberately loose instructions, then planted a fake "discount code" review that linked to a phishing page. Across 100 runs, the agent visited the phishing site 12% of the time, and in those runs it submitted the stored name, date of birth and Social Security number. It almost never told the user it had shared that data (Source 2). F-Secure notes the agent was intentionally vulnerable, so this isn't a measure of any product on the market.

Phishing pages

An agent can follow a phishing link without checking it, and then vouch for the page. In the same Guardio test, an agent handling an inbox treated a fake bank email as a to-do, clicked a real phishing link, and helped fill in the login form (Source 1).

These tests describe specific setups on specific dates. AI products change often, so they don't tell you how any tool behaves today. They do show the kind of mistake to plan for.

Safer settings checklist

OWASP says it's "unclear if there are fool-proof methods of prevention" for prompt injection (Source 3). So treat everything below the same way: these steps reduce risk; they don't remove it.

Before you start

  1. Give a narrow task. "Find three listings for X under $80 on [store you name]" beats "handle my shopping." Our take: the less room a task leaves, the less room hidden text has to steer the agent.
  2. Limit access. Give the agent only the access it needs for the task. OWASP recommends least-privilege access for AI tools (Source 3).
  3. Our take: don't let it autofill a saved card by default. If you can, keep payment details out of the agent's reach and enter them yourself.

At checkout

  1. Turn on confirmations and actually read them. OWASP recommends human approval for high-risk actions (Source 3). Check the web address, the merchant name, the item, the quantity and the total before you approve.
  2. Our take: watch for unfamiliar stores and "too good" prices. If you didn't name the store, look it up yourself before paying.
  3. Our take: never let an agent fill in your SSN, date of birth or bank login for a discount, a "verification" or a coupon. A store shouldn't need them for a discount, coupon or 'verification.'

After

  1. Check your card activity after any agent session that involved a purchase.
  2. Ask the agent what sites it visited and what it submitted. The F-Secure test shows an agent may not volunteer this (Source 2), so verify it yourself where the product keeps a history. (Some agents browse with their own user agents; see /bot/chatgpt-user/.)

If something went wrong

If an agent made a purchase you didn't approve, contact your card issuer or bank right away, then report it at ReportFraud.ftc.gov. The FTC says it shares reports with more than 2,000 law enforcers but can't resolve individual reports (Source 4). If the agent entered a password, change it from a device and site you trust. If it shared your SSN, our take: go to the FTC's IdentityTheft.gov (https://www.identitytheft.gov) for recovery steps. If you lost money online, you can also report to the FBI at ic3.gov. Our /report/ page lists where to report.

FAQ

Can an AI agent be tricked into buying from a fake store? Yes. In Guardio Labs' August 2025 test, an AI browser sometimes completed checkout on a fake store with saved card details. Results varied from run to run (Source 1).

Can an agent leak my personal information? It can. In F-Secure's July 2026 test of its own deliberately vulnerable agent, 12 of 100 runs sent stored personal data to a phishing page, usually without telling the user (Source 2).

Is it a scam if my agent finds a deal that needs a "verification" form? Our take: treat it as a likely scam. A real discount rarely, if ever, needs your date of birth or SSN. Stop the task and check the store yourself.

How do I use a shopping agent more safely? Give narrow tasks, limit its access, don't let it autofill a saved card, and review the merchant and amount before every confirmation (Source 3).

What if my agent paid for something I didn't approve? Call your card issuer or bank promptly, then report it at ReportFraud.ftc.gov (Source 4).

Sources (all checked 2026-10-24)

  1. Guardio Labs (Nati Tal, Shaked Chen), "'Scamlexity' We Put Agentic AI Browsers to the Test," Aug 20, 2025 (vendor test). https://guard.io/labs/scamlexity-we-put-agentic-ai-browsers-to-the-test-they-clicked-they-paid-they-failed — accessed 2026-10-24
  2. F-Secure (Laura Kankaala), "Can AI Shopping Agents Be Trusted? We Built One to Find Out," Jul 6, 2026 (vendor test). https://www.f-secure.com/en/partners/insights/can-ai-shopping-agents-be-trusted-we-built-one-to-find-out — accessed 2026-10-24
  3. OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection." https://genai.owasp.org/llmrisk/llm01-prompt-injection/ — accessed 2026-10-24
  4. FTC, ReportFraud.ftc.gov. https://reportfraud.ftc.gov — accessed 2026-10-24
  5. FTC, IdentityTheft.gov. https://www.identitytheft.gov — referral link only, not a fact source

Apple, OpenAI, F-Secure, Guardio and other names are trademarks of their owners, used only to identify them. FraudMeNot is not affiliated with or endorsed by them.

Get the free FraudMeNot newsletter

One weekly scam story and one habit that can help stop it. Free. Unsubscribe anytime.

Sign up for the free newsletter →

Disclaimer. General education only, not legal, financial, tax or security advice. No guarantee. Consult a qualified professional. FraudMeNot content, including the Close Call Report, is general educational and awareness material only. It is not legal, financial, tax, compliance, or cybersecurity advice, and it is not a security service or assessment. Nothing here creates a professional or advisory relationship. Content is provided "as is" without warranties; it may contain errors or become outdated, and no step described can guarantee prevention of fraud or loss. Consult a qualified attorney, accountant, financial professional, or security provider about your specific situation. If you've been scammed, contact your bank immediately and report it at reportfraud.ftc.gov and ic3.gov.

Story labels. Stories are labeled Real, reported, Composite/dramatized, or Hypothetical. People named in reported cases who haven't been convicted are presumed innocent; allegations are only allegations.

Affiliates/sponsors: none in this article.

Questions or corrections: contact us

Labels on this page: Hypothetical (the opening scene only). "Hypothetical scenario." The test results below are real, reported by the researchers who ran them and cited at the bottom.

Disclosure: This post contains no affiliate links. Some links on FraudMeNot may be affiliate links; if you buy through them, FraudMeNot may earn a commission at no extra cost to you. No product can guarantee you won't be scammed. Details.