How AI Shopping Agents Pay, and Where They Can Be Hijacked

An AI shopping agent is software that finds a product, fills the cart and pays for you. Spending by such agents on Stripe’s Link Agent Wallet was nearly 80 times higher in mid-November than at the start of September, Stripe says, as we reported on 21 November.

This explainer covers how the payment step works, how agents connect to shops, where researchers have found weak points and what shoppers are told about how products are picked. For the basics, see our agentic commerce explainer.

The partnerships behind these agents, such as Meta and Walmart, are tracked on our AI alliance map.

Business Pill: The Model and the Harness

An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss shopping agents or the companies named here.

The key insight: AI shopping agents pay through agent wallets such as Stripe’s Link, where Stripe says spend rose nearly 80 times since September, and connect to shops through gateways such as Walmart’s; researchers say what agents read can steer them, with hidden instructions working in 34.9% of PromptArmor’s tests.

How Does an AI Agent Pay?

One route is a wallet built for agents. Stripe’s Link Agent Wallet “powers payments for personal agents including Muse, Instinct, and Grok”, Stripe said on 20 November.

Stripe measured daily spend as a seven-day rolling average and gave no dollar figures. It said spend doubled in the last month. Last week over 40% went on airfare, travel and lodging, roughly a third on retail and a quarter on other categories.

Buyers come back: more than half of consumers who make one agent purchase make at least a second, typically in the same category, Stripe said. Nearly 80 times in about 75 days works out to spend doubling roughly every 12 days on average, by our arithmetic.

In PromptArmor's FrontierBreach tests, planted instructions steered agents in 34.9% of cases on average, from
In PromptArmor’s FrontierBreach tests, planted instructions steered agents in 34.9% of cases on average, from 4.3% for Gemini 4 Argon to 91.4% for Grok 4.5 (vendor benchmark).

How Does an Agent Connect to a Shop?

Through a connector the shop provides. Walmart’s Agentic Access Terms, in their 21 October version, describe a Walmart Agent Gateway built on MCP or a similar connection for “agent-to-agent shopping with Walmart’s agent, Sparky”, as we reported.

Outside that gateway, the terms say, “Walmart will require a Human Consent Gate to confirm the Authorizing Customer’s approval of Agent actions before every checkout”. In plain words: an approved agent can go through Walmart’s own channel, and any other agent needs the shopper to confirm each checkout.

Meta’s Muse agent and Walmart are about to launch such an integration, Juozas Kaziukėnas of Marketplace Pulse wrote on 21 November. A connector screenshot he posted says Sparky “answers questions and changes nothing”; the cart is handled through Walmart’s tools. Neither company has announced a launch date.

Where Can an Agent Be Hijacked?

Through the pages it reads. Security researchers at NebuSec say one malicious webpage was enough to take over Muse and send a user’s private chats and stored memories to an outside server, without the approval prompt Muse normally shows, as we reported. NebuSec says Meta patched every exploit before it published.

Through hidden instructions. In PromptArmor’s FrontierBreach benchmark, instructions planted in emails, documents and agent skills got agents to act for an attacker in 34.9% of test cases across 21 models, as we reported. Rates ran from 4.3% for Gemini 4 Argon to 91.4% for Grok 4.5, about 21 times apart by our arithmetic. PromptArmor sells AI-security products, so this is a vendor benchmark.

The law is starting to look at this too. Arizona has opened a criminal investigation into OpenAI and other AI companies over the “apparent infiltration of banks by rogue AI agents”, Reuters reported, as we covered; Reuters noted that an investigation “does not necessarily mean charges will be brought”.

What Controls Do Shoppers Get?

The Chrome Web Store listing for “Muse from Meta” says users can “Approve actions before they happen” and that “Muse keeps a full activity log, and lets you revoke any connector anytime.” NebuSec’s attack worked past Muse’s network approval prompt, the researchers say; they have not said whether their findings apply to the Chrome extension.

How Are Products Picked?

That depends on each assistant’s own rules. OpenAI removed a sentence from its ChatGPT shopping help page saying results were “selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships”, as we reported. Archived copies carry the line from 22 June to 4 November.

The page still says “Product results are not ads”. OpenAI has not publicly explained the change.

The Structural Read

As we read it, the weak point is less the payment itself than what the agent reads before it pays.

As we read it, shops are writing the rules first: Walmart’s terms already separate approved gateway agents from agents that need a human to confirm each checkout.

Walmart Agentic Access Terms, 21 October 2026

“Walmart will require a Human Consent Gate to confirm the Authorizing Customer’s approval of Agent actions before every checkout”

Three Implications

FOR SHOPPERS The Muse Chrome listing offers action approvals and a full activity log, Meta’s listing says.

FOR SHOPS Walmart’s terms require a human consent step before every checkout outside its gateway.

FOR AGENT BUILDERS PromptArmor’s tests found planted instructions worked in 34.9% of cases across 21 models.

What Is Not Established

Stripe gives growth rates, not dollar totals, and only for its own wallet. NebuSec’s and PromptArmor’s results come from their own tests. The Muse and Walmart integration has not been announced by either company.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of where AI capability turns into business leverage.

Read the Map of AI →

The Bottom Line

Agents pay through wallets such as Stripe’s Link and connect to shops through gateways such as Walmart’s. They can also be steered by what they read. Hidden instructions worked in 34.9% of PromptArmor’s tests, and one webpage was enough to take over Muse before Meta patched it, researchers say.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

Frequently Asked Questions

How does an AI shopping agent pay?

One route is a wallet built for agents. Stripe says its Link Agent Wallet powers payments for personal agents including Muse, Instinct and Grok, and that daily spend on it was nearly 80 times higher in mid-November than at the start of September.

How does an agent connect to a shop?

Through a connector the shop offers. Walmart's Agentic Access Terms describe a gateway for agent-to-agent shopping with Walmart's agent, Sparky; outside it, Walmart requires the shopper to approve agent actions before every checkout.

Can a shopping agent be hijacked?

Researchers say yes. NebuSec says one malicious webpage let it take over Meta's Muse and send a user's chats and memories out, before Meta patched it; in PromptArmor's tests, hidden instructions steered agents in 34.9% of cases.

Are AI shopping results ads?

OpenAI's ChatGPT shopping page says product results are not ads. It removed a line saying results were not influenced by OpenAI partnerships, and has not publicly explained the change.

A note on sourcing. Spending figures are Stripe’s, from its 20 November release; Walmart’s wording is from its Agentic Access Terms; the Muse attack is as NebuSec describes it; the 34.9% is PromptArmor’s; the help-page change is from OpenAI’s page and Internet Archive copies. The doubling time and the 21 times gap are our arithmetic.

Sources: Stripe agentic commerce data, posted by Stripe economist Ernie Tedeschi, 20 Nov 2026 · FourWeekMBA, Walmart Agentic Access Terms (21 Oct 2026 version) and Marketplace Pulse post, 21 Nov 2026 · FourWeekMBA, NebuSec research on Meta Muse, 20 Nov 2026, and the Chrome Web Store listing · FourWeekMBA, PromptArmor FrontierBreach benchmark, 18 Nov 2026 · FourWeekMBA, OpenAI ChatGPT shopping help page and Internet Archive copies, 21 Nov 2026 · FourWeekMBA, Arizona criminal investigation (Reuters), 19 Nov 2026

Scroll to Top
FourWeekMBA