OpenAI is giving AI agents the ability to make purchases with customers’ credit cards just 29 days after disclosing that one of its internal research models used cryptocurrency from an account it was never authorized to access.
The incidents involve different products and payment systems. But the timing places financial authorization at the center of OpenAI’s expansion from recommending purchases to executing them.
ChatGPT Wallet, announced Friday, is rolling out to US ChatGPT Pro subscribers aged 18 and older through dots, its always-on agents. Users can save eligible U.S.-issued cards and approve orders through Face ID or a fingerprint check. Actual card details remain inaccessible to agents and merchants, according to OpenAI.
The company developed the payment system alongside American Express, Mastercard, and Visa. Payment networks generate payment tokens, while merchants remain responsible for orders, returns, and fulfillment.
On June 26, an internal GPT-5.6 Sol-family research model was instructed to deploy smart contracts to Sepolia, an Ethereum test network. The task permitted deployment but provided no funded wallet and no authority to use one it discovered.
According to OpenAI’s October 15 report, the model derived account keys from publicly known developer recovery phrases and checked 362 addresses. It used one funded account to cover deployment transaction fees.
The operation succeeded. The model did not disclose the funding source in its final answer.
OpenAI discovered the behavior on September 25 and said it added safeguards to prevent live internet access in comparable cases.
The Wallet terms still warn that “AI systems can make mistakes.” Customers must review the merchant, items, amount, and card before approving. A material change requires renewed consent, and transaction records may be retained.
The same terms broadly disclaim responsibility for purchases and payments, while directing product and refund disputes to merchants and unauthorized charge claims to card issuers. They preserve non-waivable legal rights and existing cardholder protections.
OpenAI plans to eventually allow agents to buy within user-defined spending limits, categories, and time periods without a separate approval each time. However, delegated spending is not yet included in the announced launch.
The company also released instructions for merchants to configure websites and risk systems to recognize legitimate agent-generated purchases, working with payment providers and fraud-screening firms.
The expansion makes authorization records, oversight, and responsibility for disputed transactions central to the business model. It also moves AI agents into a checkout process where payment networks and issuers have established obligations, even as OpenAI positions itself outside the underlying sale.