JP Morgan and Amazon Slam the Door Shut

"Read" Allowed, "Write" Blocked

Left Unchecked, Consumers Bear the Losses

Korea Ahead on MyData

A Chance to Set the Global Standard

[Park Jongseong's AI Frontier] AI Agents Stopped at Banks' Doors... Rules, Not Keys, Are Needed View original image


JPMorgan Chase, the largest bank in the United States, has barred external artificial intelligence (AI) agents. Chief Executive Officer Jamie Dimon said on October 13, during the bank's third-quarter earnings presentation, "We will not allow external AI agents to conduct account transactions on behalf of customers." The bank will allow "read" actions, such as checking balances through intermediaries, but block "write" actions, such as transferring money and paying bills. Dimon warned that the number of agents outside the bank's control could grow to thousands, and stressed that when customers hand authority to AI, responsibility goes with it.


The issue arose with Muse, Meta's personal AI agent launched in the United States on September 8. Within 10 days of its launch, Muse overtook ChatGPT to become the top free app in the U.S. App Store. But on September 20, Amazon blocked Muse from its marketplace, saying it accessed customer accounts without prior notice and without disclosing that it was an AI. Security flaws also emerged in quick succession. A vulnerability in the Mac version of the Muse app could allow an attacker who had installed malicious code on a device to steal authentication credentials and operate the agent with the user's permissions.


The heart of the problem is market disorder caused by the absence of rules. There are no common standards defining how far an AI agent can act on behalf of a customer or who is responsible if something goes wrong. As a result, companies draw the line differently. Amazon blocked Muse entirely, while JPMorgan allowed read-only access. U.S. electronic funds transfer rules treat transactions made by a third party to whom a customer has given their password as transactions authorized by the customer, but they do not specify how to treat transactions that exceed the scope of authority delegated to an AI.


As a result, consumers have no choice but to use AI while bearing the risks, developers do not know what standards to build to, and financial institutions decide that blocking access is the safest option.


AI neon sign. Reuters-Yonhap News

AI neon sign. Reuters-Yonhap News

View original image

Blocking access may be a reasonable response to the confusion, but it is not a solution. Clear rules are needed. A financial data dispute in the United States shows just how costly a regulatory vacuum can be. In October 2024, the Consumer Financial Protection Bureau (CFPB) issued a rule requiring data sharing and prohibiting access fees, prompting banks to file a lawsuit soon afterward. The rule was subsequently sent for review, and a court halted its enforcement in October 2025. In the meantime, in July 2025, JPMorgan announced that it would charge fintech company Plaid for access to data. The industry immediately objected after reports suggested that Plaid could have to pay $300 million a year under the initial fee schedule.


Nevertheless, JPMorgan signed an agreement with Plaid in September 2026, with the terms kept confidential. The costs were bound to be passed entirely on to consumers. During the two years in which public standards remained stalled, companies with greater bargaining power controlled how the market operated, and no one was there to represent consumers' interests.


Signs point to the same thing happening in the AI agent market. On October 6, Meta and AI startup Sierra announced an open standard called the Personal Agent Protocol (PAP). An agent first accesses public information without logging in. If it needs access to a customer account for tasks such as changing a reservation, the customer logs in directly through the company's interface and grants either read or write permissions. Walmart, Stripe, and Shopify are among the participants, but payment functionality was left out of the initial specification.


Fortunately, South Korea is somewhat ahead in this area. It has already addressed the question of read access that is still being contested in the United States. When the MyData system, formally known as the personal credit information management business, was fully implemented in January 2022, scraping - the practice of using customers' IDs and passwords to retrieve information from financial institutions' interfaces - was prohibited. Data could thereafter be exchanged only through standardized APIs. The key change was who transferred the information. Instead of companies using customers' keys to open the door and retrieve data, financial institutions now send only the information customers have authorized. This has curbed indiscriminate data collection and clarified who is responsible in the event of a leak. A regulatory exemption has also laid the groundwork for write actions. The "interest rate reduction request service" uses AI to detect improvements in a customer's credit profile and request a rate cut from a financial institution on the customer's behalf. Since being designated an Innovative Financial Service in December 2025, the service had received applications from about 4.07 million people by July 2026, helping 162,000 people reduce their total interest burden by 100.3 billion won.


What is needed now is the finishing touch. On August 25, the Financial Services Commission outlined a direction for amending the Credit Information Act that would expand the scope of delegated actions to include applications for policy loans and refinancing, and allow advance blanket consent specifying the purpose, scope, method, and duration. It also said that subordinate legislation would set out requirements to record delegated actions and explain their outcomes. But to open the door to full-fledged write actions, such as transfers and payments, where a mistake can immediately drain money from an account, guidelines must address three things: delegation standards that set limits by amount, number of transactions, and duration; criteria for sharing losses if something goes wrong; and a common authentication standard requiring external agents to identify themselves as AI, prove the scope of their delegated authority, and gain access.


Speed matters, too. Since amending the law will take time, pilot programs for Innovative Financial Services should be run alongside updates to the guidelines to narrow the regulatory gap. While the United States is still debating the standards, South Korea can complete the rules for write actions first and smoothly test the safety of financial AI agents. This could become the strongest advantage for South Korean financial institutions and fintech companies as they compete globally.



Park Jongseong, writer and AI consulting lead


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing