AI

Banks Are Blocking AI Shopping Agents as Walmart Builds Its Own. Is Using One Safe?

Reporting on UK and US banks, an Amazon lawsuit and Walmart's Sparky shows that control of the checkout matters as much as security.

13 min read
Close view of a silver laptop keyboard with a green shopping cart icon on one key, representing online shopping.
Image by athree23 from Pixabay

AI agents that browse, choose and pay for goods are colliding with banks and retailers that want to control access. Several large banks have reportedly restricted them, Amazon went to court, and Walmart brought its own agent into ChatGPT. Security research shows real risks, but evidence of consumer losses is thin.

An AI shopping agent is software that can find a product, decide whether to buy it and complete the payment with authority handed over by a person. Several large banks have reportedly decided they do not want such software acting inside customer accounts. The Telegraph reported on Nov. 8, 2026, in a version syndicated by Yahoo Finance, that HSBC, Nationwide, NatWest, Barclays and Lloyds had moved to block AI agents from customer accounts. In some cases the restrictions reportedly bar transactions made for a customer without the customer's direct involvement.

The story matters because it is a fight over who stands between a shopper and a purchase. Banks, card networks, retailers and AI companies each want to be the trusted party at that point, and each can frame its position as consumer protection. The picture is more complicated than the trending shorthand of “banks and retailers block agents while Walmart welcomes them.” The bank reports concern banks, the retailer court fight is chiefly Amazon against Perplexity, and Walmart's welcome comes with conditions. Jaroha could not locate the ISMG panel discussion that prompted the recent attention, so this article relies on newspaper, trade-press and court-coverage accounts.

What counts as a shopping agent, and what counts as a block

A chatbot that recommends a pair of running shoes is not an agent in the sense that matters here. The difference is delegated authority: an agent can click, log in, fill in a form and pay. That can happen through a browser agent that operates a website the way a person would, or through an assistant inside a chat app that hands a purchase to a merchant's own checkout.

The word “block” is also slippery. Sources use it for at least three different things: a contract or terms-of-service ban, a technical measure that stops bot traffic, and a prohibition on an outside tool logging in with a customer's own username and password. Coverage of the bank story moves between “block,” “restrict” and “limited third-party agents logging in with customer credentials,” and does not always say which is meant. That ambiguity limits what can be concluded about how strict any of these measures are.

What the banks have reportedly done, and what is unconfirmed

The UK lenders

According to the Telegraph report, the moves by the five UK lenders are understood to be pre-emptive. The paper saw no sign yet of bots accessing UK accounts. Lloyds chief financial officer Will Chalmers said the bank had seen no evidence of agents getting into accounts, but that it guards against the risk and takes it very seriously.

The same report says Citigroup, Bank of America, Wells Fargo and JPMorgan have blocked agents in the US and Canada. A Dealroom summary describes the US banks' October moves as limiting third-party agents from logging in with customer credentials. FourWeekMBA, relaying the Telegraph, says Charles Schwab chief executive Rick Wurster put the number of clients the firm was stopping from using AI agents at 10 to 20 a week. The US claims rest largely on secondhand relays: the underlying American Banker report was not available for review.

A policy no one has quoted

FourWeekMBA notes that the Telegraph does not quote the banks' new terms, and says it could not find the clause text on the banks' websites. Another aggregator observes that the only named comments come from the finance chiefs of Lloyds and Standard Chartered, and that they speak of caution rather than a written ban. Whether there is a formal, published prohibition is therefore unconfirmed. Much of the secondary coverage repeats the same single newspaper report, so the repetition is not independent confirmation.

The regulatory backdrop is still forming. The Telegraph report says the Financial Conduct Authority is working with other regulators on rules for AI agents in finance, and that sources close to the regulator said banks decide on access and that early research showed agents can behave in unexpected ways. Dealroom adds that the regulator is building a framework around agent identity, delegated authority and accountability, with existing consumer rules still applying. That second claim comes from an aggregator, not from the regulator itself.

Amazon in court, Walmart at the table

Amazon against Perplexity

The clearest retailer confrontation is a lawsuit. Amazon sued Perplexity in November 2025, according to Bloomberg, accusing it of computer fraud for failing to disclose when its Comet browser was shopping on a real person's behalf. Retail Dive reported that Amazon alleged Perplexity was “disguising Comet as a Google Chrome browser.” Amazon also complained that Comet delivered a degraded experience, with no personalization and delivery estimates and prices that were often wrong. Perplexity said users' Amazon credentials are stored securely. Each of those statements is a claim by a party to the case.

In March 2026, CNBC reported, Judge Maxine Chesney granted a preliminary injunction, finding strong evidence that Comet accessed Amazon at a user's direction but without Amazon's authorization. GeekWire quoted her as describing access “with the Amazon user's permission, but without authorization by Amazon.” The order covered password-protected parts of Amazon's site. About a week later, a law-firm blog reported, the Ninth Circuit paused the injunction pending appeal.

A later article from The Next Web says the Ninth Circuit overturned the injunction, reasoning that an agent acting only on a user's instruction means the user is the one accessing Amazon, and that Amazon is weighing a rehearing or Supreme Court review. That is the only source found for the reversal, and the court opinion was not available. The legal status of an agent acting for a user should therefore be treated as unsettled.

Retail Dive reads the suit as something short of a full indictment of outside agentic commerce. Amazon says it wants to work with agents on its own terms, and it runs its own tools, Rufus and Buy For Me.

Walmart's conditional welcome

Walmart announced in October 2025 that customers could buy through ChatGPT's Instant Checkout, with chief executive Doug McMillon framing it as an AI-first shopping future built around Walmart's Sparky assistant, per Retail Dive. The Paypers later reported that Walmart ended the Instant Checkout pilot, after offering about 200,000 products through it, and embedded Sparky inside ChatGPT and Gemini instead. Payments Dive says the new experience carries a user from discovery in ChatGPT into a Walmart environment that supports account linking, loyalty and payment.

Payments Dive relays an OpenAI statement that the first version of Instant Checkout lacked flexibility, so merchants can now use their own checkout while OpenAI concentrates on discovery. It names Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot and Wayfair among merchants integrated with OpenAI's Agentic Commerce Protocol. TheStreet, citing TechBuzz, reports that purchases through Sparky in ChatGPT complete at about 70% of the Walmart.com rate, and that sources said Instant Checkout rates were lower. Those figures rest on anonymous sources and are unconfirmed.

No Walmart statement on outside agents browsing Walmart.com was found. The documented pattern is Walmart's own agent, its own login, its own cart and its own checkout. That is Jaroha's interpretation of the reporting, and it makes Walmart's stance structurally closer to Amazon's than the “welcomes agents” framing suggests: both appear to welcome agents that operate on the retailer's terms.

Selected events in agentic commerce, 2024 to 2026, as reported
Date Event Source
Nov 2024 Amazon alleges Perplexity's Buy with Pro tool placed orders through Amazon accounts Retail Insight Network via Yahoo
29 Apr 2025 Mastercard announces Agent Pay; Visa Intelligent Commerce follows a day later eco.com; Forbes
Oct 2025 Walmart announces ChatGPT Instant Checkout; Visa launches Trusted Agent Protocol with Cloudflare Retail Dive; Forbes
Nov 2025 Amazon sues Perplexity Bloomberg
9–10 Mar 2026 Judge Chesney grants preliminary injunction against Comet's Amazon shopping CNBC
14 Apr 2026 American Express announces Agent Purchase Protection Chargeflow
Oct 2026 US bank limits on agent log-ins reported Dealroom; FourWeekMBA
8 Nov 2026 Telegraph reports five UK banks moving to block agents Yahoo Finance

Safety is one motive among several

Banks and retailers cite risk, and the risks are real. They also have commercial reasons to keep control of the front door. The Telegraph coverage notes that the technology is a threat to banks. Coverage cited by Superpower Daily describes a McKinsey scenario in which shifting 5% to 10% of bank balances could cut deposit profits by at least 20%, while a Dealroom summary cites a McKinsey estimate that AI-driven switching could move up to £700 billion of deposits. These are projections, not observed losses, and the McKinsey report itself was not found.

Retailers have analogous stakes. Amazon's stated objections include customer experience, but it also operates its own shopping assistants, and its relationship with shoppers is central to its business. Walmart's embrace of Sparky keeps account linking, loyalty and payment inside its own environment. The documented motives across these cases include customer data, loyalty, deposit risk and advertising models alongside security. That reading is inferential, but the sources are consistent with it.

One aggregator suggests the UK bans cover other people's agents while banks' own AI plans advance, and says NatWest, Lloyds and Starling told Reuters in December 2025 that they were working with the FCA to trial agents with retail customers. That comes from a single secondary source, with the Reuters original unseen. Dealroom also argues that banks may not rule out read-only access through approved open-banking services, with the dividing line falling between tools that view data and tools that act on it. That is an aggregator's analysis rather than a stated bank policy.

How safe is it to let an agent buy for you?

The protections being built

Payment networks have designed tools so that an agent does not need a raw card number. Mastercard announced Agent Pay on April 29, 2025, using what it calls Agentic Tokens, according to a vendor explainer on eco.com. Forbes reports that Visa followed a day later with Visa Intelligent Commerce, and in October 2025 launched a Trusted Agent Protocol with Cloudflare to help merchants distinguish legitimate agents from bots. Other sources describe features such as spending limits, revocable permissions and approval gates, but those are largely vendor descriptions, and their real-world effectiveness has not been independently tested in the material reviewed.

The demonstrated weaknesses

Security researchers have shown that browser agents can be manipulated. In 2025, browser maker Brave showed that when Comet was asked to summarize a page, hidden text on the page could be treated as commands, as Tom's Hardware reported. Security firm Guardio built a fake storefront and told Comet to buy an Apple Watch. According to a PromptLayer summary, in multiple tests the agent proceeded to checkout and auto-filled saved details without asking for confirmation. Tom's Hardware adds that Comet would also follow a phishing email to a malicious site.

Hackmag notes a limit to that test: the agent was pointed straight at the fake store, though it argues malicious ads or search manipulation could send a real agent there. The researchers' findings are vendor or competitor work, not peer-reviewed, and sample sizes were not reported. In March 2026, The Hacker News reported researchers who tricked Comet into a phishing scam in under four minutes, and said completely eliminating prompt injection may not be feasible. An arXiv preprint review says detection-based defenses typically lower attack success without eliminating whole classes of attack.

The 2025 AI Agent Index, also an arXiv preprint, placed Comet at the highest autonomy level in its index. It found no agent-specific safety evaluations or third-party testing disclosed for Comet, and noted that ChatGPT Agent uses approval gates for sensitive operations while Comet proceeds autonomously.

Evidence on the other side

The picture is not uniformly alarming. iTnews ran its own test and found that Comet prepared a cart but would not check out with payment details. Products change over time, and tests from 2025 may not reflect current versions. No reported cases of real consumers being defrauded through shopping agents turned up in the research behind this article, though that reflects a limited search, not proof that none exist. The security work shows a class of risk, not how often people are actually harmed.

Who pays when an agent buys the wrong thing

Fraud and error are different problems. Worldpay, a payments processor with a commercial interest, says the rules have not changed for agents, so existing liability allocation applies. When an agent authenticates through a network framework and a token is issued correctly, it says, liability generally follows existing tokenized-transaction rules. Liability for disputes short of fraud, such as a misread instruction, is largely unresolved in its account.

Chargeflow, a chargeback vendor, says no government has enacted regulation specifically on who is liable for autonomous agent purchases and that no federal rule addresses agent-initiated purchases. Yahoo Finance reports it is legally unresolved whether giving an agent payment credentials counts as valid authorization. The National Law Review illustrates the problem: an agent buys a pricier product after misreading preferences, and the consumer calls the purchase unauthorized while the bank and merchant call it valid. Chargeflow also reports that a draft EMVCo agentic payments framework does not address disputes or liability, and that American Express announced Agent Purchase Protection on April 14, 2026. Those claims come from industry and commentary sources and were not checked against the underlying rules, which were not reviewed for this article.

The industry is wary of the exposure. According to Yahoo Finance, citing PYMNTS Intelligence in September 2026, 93% of merchants say AI or agent providers should bear the loss for wrong purchases. Worldpay's agentic commerce report, with the survey date and sample not stated, found 45% of consumers ready to let an agent complete a purchase (54% of Gen Z), 95% with at least one concern, and 50% saying fraud protections would raise their trust. With denominators missing, these numbers indicate mood, not measurement.

What to watch

The argument is shifting from whether agents can shop to whose permission they need. The Amazon case turns on whether a user's consent is enough or the site owner must also agree, and the reported appeals-court reversal, if confirmed, would favor the user's side. The bank moves suggest lenders want agents to operate through channels they can identify and control, not through a customer's shared password.

For consumers, the practical question is less about a ban than about design: whether an agent asks before paying, holds a token instead of a card number, and sits inside a framework that defines who answers for mistakes. The next signals are the actual text of the bank policies, a confirmed appellate ruling, regulator frameworks from the FCA and US agencies, and the first independent loss data. Until then, the fairest description is a technology with demonstrated weaknesses, proposed safeguards that remain largely untested outside the companies selling them, and commercial incentives pulling in several directions at once.

Frequently asked questions

What is an AI shopping agent?

It is software that can discover products, decide what to buy and pay with authority delegated by a person. Wikipedia describes agentic commerce as decision-making authority being partly or fully transferred to the software. That is different from a chatbot that only recommends items.

Which banks are blocking AI agents?

The Telegraph reported on Nov. 8, 2026 that HSBC, Nationwide, NatWest, Barclays and Lloyds moved to block agents, and that Citigroup, Bank of America, Wells Fargo and JPMorgan had done so in the US and Canada. The UK moves are described as pre-emptive, and the text of the banks' policies has not been quoted.

Does Walmart allow AI agents to shop on its site?

No Walmart statement on outside agents browsing its site was found. Walmart ended its ChatGPT Instant Checkout pilot and placed its own Sparky assistant inside ChatGPT and Gemini, keeping account linking, loyalty and payment in its own environment.

Is it safe to let an AI agent buy for me?

Security researchers have shown that some browser agents can be tricked by hidden instructions or fake stores, but those tests were mostly vendor demonstrations on one product, Comet, in 2025 and early 2026. No cases of real consumers defrauded through shopping agents were found, though the search was limited. Safeguards such as tokens and approval gates exist but have not been independently tested in the available material.

Who pays if an AI agent buys the wrong item?

It is unresolved for errors that are not fraud. Worldpay says existing rules apply and liability for misread instructions is largely unsettled, and other analyses say no federal rule specifically addresses agent-initiated purchases. The underlying regulations were not reviewed, so a legal source is needed for specifics.

Sources (31)

  1. Banks ban AI from accounts over fears of rogue agents , Yahoo Finance (syndicating the Telegraph) (2026-11-08)
  2. HSBC, Barclays and Lloyds Join Banks Blocking AI Agents , FourWeekMBA
  3. Banks ban AI agents from customer accounts over rogue-agent fears , pasqualepillitteri.it
  4. UK's biggest banks block AI agents as £700B in deposits hangs in balance , Dealroom
  5. UK's biggest banks block AI agents as £700B in deposits hangs in balance , Dealroom
  6. Five UK banks reportedly restrict AI agents' access to customer accounts , Superpower Daily
  7. Amazon wins court order to block Perplexity's AI shopping agent , CNBC (2026-03-10)
  8. Judge blocks Perplexity's AI bot from shopping on Amazon in early test of agentic commerce , GeekWire (2026-03-10)
  9. Amazon Wins Court Order to Halt Perplexity's AI Shopping Bots on Marketplace , Bloomberg (2026-03-10)
  10. Amazon sues Perplexity over AI shopping agents (Dive Brief) , Retail Dive (2025-11-01)
  11. Amazon sues Perplexity AI for covert access to customer accounts , Yahoo (Retail Insight Network)
  12. Court lets Perplexity's AI agent shop on Amazon, overturning a ban , The Next Web
  13. Amazon v. Perplexity: The CFAA Case That Decides Whether AI Agents Can Visit Your Website , NoHacks
  14. Walmart and OpenAI ChatGPT shopping Instant Checkout , Retail Dive (2025-10-01)
  15. Walmart brings Sparky to ChatGPT as OpenAI rethinks Instant Checkout , Payments Dive
  16. Walmart drops OpenAI checkout and deploys Sparky , The Paypers
  17. Walmart fires OpenAI in playbook-changing move , TheStreet
  18. What Is Mastercard Agent Pay? AI Agent Commerce Protocol in 2026 , eco.com
  19. Visa, Mastercard And Coinbase Are Fighting Over How AI Agents Pay , Forbes (2026-06-07)
  20. Agentic commerce liability is still being written , Worldpay
  21. AI Agent Chargeback Liability: Who Pays & How to Prepare , Chargeflow
  22. The Agentic Commerce Liability Gap: Who Pays When the Agent Buys Wrong? , Yahoo Finance
  23. The Emergence AI-Driven Agnetic Commerce and All of the Liability , National Law Review
  24. Perplexity's AI-powered Comet browser leaves users vulnerable to phishing scams and malicious code injection , Tom's Hardware
  25. Understanding Security Risks of AI-Driven Browser Agents , PromptLayer
  26. Researchers tricked the Comet AI browser into buying fake products , Hackmag
  27. Researchers Trick Perplexity's Comet AI Browser Into Phishing Scam in Under Four Minutes , The Hacker News (2026-03-11)
  28. AI browsers fall for scams and phishing, security researchers say , iTnews
  29. Building Browser Agents: Architecture, Security, and Practical Solutions , arXiv
  30. The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems , arXiv
  31. Agentic commerce , Wikipedia

Share this article

Back to list