Act this weekWhat happenedA security article warns that AI shopping agents should not receive the one-time codes retailers send to confirm a buyer's identity.
Who is affectedAnyone who shops online or lets AI assistants buy for them, and the retailers and payment companies building these tools.
What to do- 1Never paste or read a security code to an AI assistant, chatbot or anyone who asks.
- 2Approve purchases yourself on your own phone, after checking the item, price and merchant.
- 3Set spending limits and allowed stores for any AI agent you let buy for you.
- 4Review and remove the permissions you gave an AI shopping tool if you no longer use it.
AI shopping assistants can now find a product, compare prices and reach checkout for you. Then the store texts a one-time code to confirm it is really you. A security article says the code should never go to the assistant. You approve the purchase yourself, on your own phone.
What changed
Visa, Mastercard, Google, OpenAI and Stripe have all introduced ways for AI agents to start payments on a customer's behalf. How those agents handle the extra identity check is being decided now. The easy answer is to let the customer pass the code to the agent. The article argues that is a mistake, because it turns a security step into one more box to tick.
Why a used-once code is not enough
A one-time code works only once, but that does not make it scam-proof. NIST, the US standards body, treats these codes as weaker than phishing-resistant methods, because an impostor can type in a code that was sent to you. Adding an assistant adds another place the code could leak, such as chat history, connected tools or a compromised workflow.
There is also a people problem. If real assistants ask for codes, scammers get a believable line: your shopping assistant needs this code to finish your order.
Who can relax
The article says this is a foreseeable risk, not proof that a wave of AI-related code fraud has begun. Shoppers who never let an assistant buy for them are not in new danger. Codes also still work well when they go straight to the person they were meant for.
What to do
The steps in the alert box above cover the basics: keep codes to yourself, approve on your own phone and limit what any agent can spend.
What happens next
The article calls on retailers and payment companies to build in a clear line between what an agent can start and what a customer has approved. It describes what good approval would look like:
- A request sent straight to your phone, never through the agent
- A message showing the items, total price, merchant and payment method
- Approval tied to that amount and that seller, so any change needs a new yes
- Spending limits and permissions set by the payment system, not just the agent's instructions
EU payment rules already work this way. An approval code there is tied to one amount and one payee, and changing either makes it fail. Bank of America has also published principles for agentic commerce. How soon stores and payment firms adopt such protections is not yet known.
The takeawayIf any app, assistant or person asks for a security code, say no and approve the purchase yourself on your own phone.