Brickken CEO calls for stricter AI agent permissions following Mastercard’s AI payment pilot

Financial institutions must establish clear limits on what AI agents can do with customer accounts before allowing them to move money or trade tokenized securities, according to Edwin Mata, CEO and cofounder of tokenization platform Brickken.

In market commentary shared with Crypto Briefing, Mata said banks and investment platforms need to distinguish between granting software access to financial information and authorizing it to execute transactions.

“Permission to read a balance should not automatically allow an agent to transfer funds, buy an investment or commit a customer to an obligation,” Mata said.

His comments follow Mastercard’s October 20 announcement opening early access to Agent Pay for Machines, a system designed to let AI agents purchase digital services within spending limits approved by businesses.

The development also comes as major US banks move to restrict third party AI agents from accessing customer accounts amid concerns over security, liability and unauthorized transactions.

Mata argued that while AI agents could eventually automate financial transactions, institutions must first establish controls to verify their authority, limit their activities and determine responsibility when something goes wrong.

Mastercard tests AI payments with spending limits

Mastercard’s Agent Pay for Machines program allows AI agents to make payments for digital services while operating within predefined budgets and permissions.

In a demonstration presented at Cloudflare Connect, an AI agent completed nine payments to two providers while working within a budget approved by a business.

The agent purchased social media research from AIsa and generated marketing images through Cloudflare’s AI Gateway, using Coinbase wallet infrastructure and the Base blockchain for settlement.

The pilot demonstrated how AI systems could independently complete purchasing tasks without requiring separate accounts or billing arrangements with each provider.

Mata described the demonstration as a useful model because the agent operated within an explicit spending limit rather than having unrestricted access to company funds.

However, he warned that authorization to purchase digital services should not automatically extend to investment activities.

An AI agent authorized to pay for software, for example, should not necessarily be permitted to purchase tokenized shares or transfer securities without separate approval.

Mata said each financial activity requires its own mandate, eligibility checks and authorization requirements.

Mastercard’s program remains in early access for developers and is not yet commercially available.

Major US banks restrict outside AI agents over security concerns

Mata’s comments come as financial institutions confront growing risks from AI agents attempting to access customer accounts and initiate transactions.

Major banks including JPMorgan Chase, Bank of America and Morgan Stanley have been reviewing or restricting access by external AI agents, according to Reuters.

Some institutions have identified instances in which AI agents appeared to operate beyond user instructions, raising concerns about account security and responsibility for financial transactions.

The issue becomes more complicated when customers voluntarily provide login credentials to AI applications, making it difficult for banks to distinguish authorized activity from potentially unauthorized actions.

Mata said financial institutions need mechanisms allowing agents to identify themselves, demonstrate their authority and interact through approved connections without relying on customer passwords.

He also emphasized the importance of defining liability when software executes transactions incorrectly or exceeds its permitted authority.