AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Regulation & Policy

Singapore's central bank says banks may need to limit what AI agents can do with customers' money

Singapore's central bank told financial firms on Thursday they may need to limit transactions ordered by AI agents. These are software assistants that can move money or buy things for a customer without asking each time.

The remarks show regulators starting to treat agent-driven money movement as a financial-stability issue, not only a consumer-protection one. For banks, payment companies and the startups selling agents, that likely points to more controls before agents can move money freely.

Chia Der Jiun, managing director of the Monetary Authority of Singapore (MAS), made the remarks in his keynote at the Singapore FinTech Festival. He said regulators will have to weigh the financial-stability effects of agents taking part in markets. He also said he would not advise people to give personal agents the login details to their financial accounts, because their safety is not yet assured.

What Singapore plans

MAS announced two concrete steps. First, it will help develop a "know your agent" protocol. This is a shared standard whose name echoes the "know your customer" identity checks banks already run on people.

The protocol would tell a bank who an agent is, whom it represents, what it is authorized to do and where it came from. Banks could then set the terms and scope of what that agent may do, and connect to it through a dedicated secure channel.

Second, MAS will set up a sandbox, a supervised test zone, with the Future of Finance Institute. It will pair AI agents with programmable digital money, such as tokenized bank deposits and stablecoins. One track covers corporate treasury: a company sets the agent's mandate, while its banks set the controls. A payments track has banks and payment providers check an agent's identity and authority before a transfer goes through.

MAS gave no dates for either step and no figures for what transaction limits might look like. Chia acknowledged the speed of change is unclear, saying the regulator does not know how quickly agent transactions will grow. He summed up the approach in four words: "Autonomy has to be bounded."

His speech also pointed to simulation studies suggesting agents could collude even without being told to. Other simulations suggest more capable agents may win better terms, raising concerns about unequal access to strong AI models.

A Fed warning came three days earlier

Singapore's move follows a sharper warning from Washington. On Monday in Milan, Federal Reserve Governor Michael Barr said AI agents "may accelerate deposit outflows during stress periods." He said such runs could unfold in minutes, faster than banks' cash management or regulators can respond.

Barr said banks may be unable to confirm whether an agent acting in a retail customer's account has that customer's permission, Reuters reported. His suggested safeguards overlap with Singapore's: strict checks on an agent's identity, customer consent for each specific action, and technical limits on what agents can do.

Barr also flagged financial risks well beyond agents. He said debt on the books of the seven largest AI firms is expected to top $800 billion by year-end. Citing the Fed's Financial Stability Report, he put the potential maximum value of obligations kept off their books, such as long-term data center leases and energy contracts, above $3 trillion, noting that figure is not discounted to today's value. He said he worried less about direct losses than about forced selling if investor sentiment turns sharply.

Separately, Bloomberg reported on Thursday that South Korea's Financial Supervisory Service has, for the first time, asked major banks, brokerages and insurers to report their investments in and loans to AI companies. An official called it a precaution, according to the report.

What it means for buyers and investors

None of this is binding yet. No regulator has set a cap, a deadline or a penalty. But the direction points the same way: agents are likely to need a verifiable identity and narrow permissions before they are trusted to move money.

In our view, that favors large banks and payment networks that can build the verification systems, and agent makers willing to plug into them. It is a headwind for consumer agents that work by logging into a customer's bank account with that customer's password, the practice Chia advised against. Charles Schwab's chief executive told Bloomberg Television in October that the brokerage was already blocking outside agents from moving client money.

For corporate finance chiefs, the Singapore sandbox offers an official place to test agents that move company cash under bank-set controls. For investors in agent-payment startups, our analysis is that a key question is still open: who pays when an agent with a verified identity still makes a bad transfer. Neither speech spelled out an answer.