Bank Runs in Minutes, Driven by AI Agents? Where the Chain Breaks in Europe

Fed Governor Michael Barr warns that AI agents could drain bank deposits within minutes. We checked each link in the chain: what agents can actually do with your account today, what US banks are already blocking, and what simulations show. We also looked at whether Europe's two-factor rule would stop a run like that.

Analysis Nov 16, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What's going on

Michael Barr is a governor of the US Federal Reserve. On November 16, 2026, he spoke at an industry conference on AI in finance in Milan. His speech contains the line that Bloomberg and other outlets have picked up since: AI agents could "instantly detect" rate differences or signs of crisis "and autonomously move money at high speed." That, he said, could speed up bank runs "that play out in minutes." AI agents are programs that don't just answer questions but take actions themselves, such as logging in, filling out forms and making payments.

The warning sounds dramatic. But an agent-driven bank run would need a whole chain of conditions to hold. Agents would need access to accounts, permission to move money without a human signing off, and they would all have to panic at the same moment. We check that chain link by link, with a focus on where it breaks in Europe.

What Barr proves and what he doesn't

Barr's speech offers just one real-world piece of evidence about the present. Large banks reported finding AI agents trying to carry out actions in retail customer accounts in ways the bank can't verify. The bank can't tell whether the agents are legitimate, whether the customer consented, or whether they're following the customer's instructions. Barr names no banks and gives no case numbers or loss figures. The speech also contains no data on deposit flows. The bank run in minutes is a warning about the future, not something anyone has observed.

As countermeasures, Barr calls for strong authentication, explicit consent for individual actions, and technical limits on what agents are allowed to do. As we show below, that's already law in the EU.

Link 1: What access agents really have today

Access Read Move money Approval
Shared password (agent logs in as the customer) yes technically yes blocked by major US banks
Meta Muse via Plaid (US) yes, read-only no consent per transfer, according to CSA
Gemini with Google Wallet (US) spending, passes, rewards no "reads and summarizes, doesn't act" (TalkAndroid)
ChatGPT Wallet (US, Pro subscription only) – card purchases Face ID or fingerprint per purchase
Scalable "Agentic Investing" (DE) portfolio, cash balance securities orders, no payments every order must be confirmed
Schwab's own MCP servers yes trading in pilot, no money transfers Schwab controls the tool list

MCP is a standard that lets a bank offer an agent a fixed list of permitted tools, a bit like a counter that only handles clearly defined services. According to Finanztip, at Scalable you confirm every order in the chat. The link itself is confirmed with two-factor approval and renewed "at recurring intervals." Payments are "not possible via the AI interface." So an agent can sell ETFs there, but it can't move the money out. ChatGPT Wallet uses single-use card numbers called tokens and requires biometric approval for every purchase. OpenAI has announced it will eventually let agents make purchases within preset limits on their own, but it hasn't given a date.

Our take: the only way money could flow without approval today is the gray route of shared passwords. And that's exactly the route banks are shutting down.

Link 2: What major US banks already block

According to Reuters, as cited by Börsen-Zeitung, Morgan Stanley blocks external AI applications. Insiders say the bank hasn't seen any cases of agents acting on their own, though. According to BERI, JPMorgan steps in when an outside agent tries to log in with the customer's banking credentials. Agents may read data there only through authorized services like Plaid, and they can't send money. Citi doesn't let outside agents into accounts at all.

Schwab CEO Rick Wurster has been the most specific. On Bloomberg TV, he said Schwab blocks agents "in the tens and twenties" every week. Speaking to American Banker, however, he called such cases "rare." BERI notes that the two statements don't add up. At Schwab, the bank's own models get access through Schwab's servers, but:

"We don't allow them to move money, because we think the risk to the client and the firm is too great."

Some background: the US has no mandatory two-factor approval like the EU's. A federal court in Kentucky halted the CFPB's open banking rule, §1033, and according to Open Banking Tracker the consumer protection agency is now rewriting it. That leaves it up to the banks to decide whether agents get into accounts.

Link 3: Would agents panic in the first place?

One of Barr's sources is the paper "Ex Machina" by Anand, Kazinnik, Leonello and Panetti (ECB Working Paper 3225). The researchers had AI investors play a fund redemption game in which they decided whether to pull their money. According to the ECB Research Bulletin, Q-learning agents, which learn by trial and error, all withdrew together, even when fundamentals were solid. Every failure they experienced in training made them jumpier. The authors call it the "hot stove effect," as in "once burned, twice shy." LLM-based investors, on the other hand, never withdrew when the data was solid. Past failures didn't affect them, because they don't learn from earlier payouts. They only became unpredictable in the gray zone of middling values.

The catch: when the language models received private signals about the situation, their expectations converged and they acted the same way. That has a stabilizing effect as long as the signals are good. If the signals are bad or manipulated, everyone runs in the same direction. A not-yet-peer-reviewed arXiv preprint by Ruano and Rajan simulates LLM-based savers in a Twitter-style network. More connectivity increased both the likelihood and the speed of withdrawal waves. Above a transmission rate of about 0.10, contagion between banks jumped sharply. So panic is neither ruled out nor a given. What matters is what all the agents are reading at the same time.

Link 4, the crux: Does PSD2 stop an agent-driven run in Europe?

In the EU, the PSD2 payment services directive mandates strong customer authentication: two independent factors, such as a password plus a smartphone. For transfers, the approval is also tied to the amount and the payee. An agent with a stolen or shared password won't get far as long as the phone stays with the customer. The back doors are in the exemptions of Delegated Regulation 2018/389. We checked them one by one:

Exemption Condition Useful for pulling out large sums?
Account information (Art. 10) balance and transactions only, re-authentication after 180 days at the latest no, read-only
Trusted beneficiaries (Art. 13) list set up by the customer with two-factor approval; banks can choose whether to offer the exemption limited: only to pre-approved destination accounts
Recurring transactions (Art. 14) same amount, same payee; setting up and changing requires approval no, the amount is fixed
Own accounts at the same bank (Art. 15) same person, same institution no, the money stays in the bank
Low-value payments (Art. 16) up to €30, max. €100 or five payments in a row no
Transaction risk analysis (Art. 18) transfers up to €500 at a fraud rate of no more than 0.005% no

So standing orders aren't a back door. The only real gap is trusted beneficiaries. If you've added your savings account at another bank as a trusted beneficiary yourself, and your bank uses this exemption, transfers to that account can go through without fresh approval. Even then, ongoing fraud monitoring still applies. And if you give an agent your password, you're usually breaching your bank's terms and may be liable yourself if something goes wrong. The UK shows it can be done differently. There, "sweeping" with variable recurring payments (VRP) allows automatic transfers between your own accounts at different banks without approval for each payment. It's built on exactly these exemptions, and providers include HSBC (Open Banking UK).

What about the upcoming EU Payment Services Regulation (PSR)? The final compromise texts have been available since April 2026. According to Avanto, the regulation hasn't been published yet and probably won't apply before late 2027 to 2028. Law firm analyses say it allows two-factor checks to be delegated to wallets and technical service providers, treating this as regulated outsourcing. A dashboard is meant to show which third parties you've given access to. We found no specific rules for AI agents in the available summaries. In its September monthly report, the Bundesbank considers it an open question whether an agent mandate counts as consent under Article 64 of PSD2. It floats the idea of "mandatory human review for large transactions."

Our take: the real back door isn't the exemptions but future mandates. That means the "agent can spend up to X euros on its own" model that OpenAI has announced. Whether that will be allowed in the EU, and with what limits, is still unclear.

The benchmark: SVB managed it without agents

On March 9, 2023, Silicon Valley Bank customers withdrew about $42 billion, roughly a quarter of its $175 billion in deposits. Another roughly $100 billion was queued up for the following day. Barr himself told the Senate in 2023 that the run happened "in less than 24 hours." Spread $42 billion over 24 hours and you get an average of about $29 million per minute (our own calculation), all triggered by humans using banking apps. Notably, a new Fed review under Vice Chair Michelle Bowman concluded in September 2026 that social media neither triggered nor accelerated the run, according to PYMNTS. It found that 96 percent of the discussion only appeared once the collapse was already inevitable. The decisive factors were that 94 percent of deposits were uninsured and that losses on securities exceeded the bank's equity.

You can also run the numbers on the rate gap that Apollo chief economist Torsten Sløk cited in September when he warned of an "agentic bank run." With 0.1 percent on a US checking account versus 3.3 to 5 percent at fintechs, a saver with $10,000 misses out on $320 to $490 a year. That's a reason to move money gradually, not a reason to panic. According to American Banker, experts consider an agent-driven run "possible but unlikely."

The bottom line

Barr's warning is a fair prompt for thought. When many similar agents read the same signals, they can act in lockstep, and the simulations bear that out. But today the chain leading to a bank run in minutes is broken in several places. No major agent service can withdraw money without human approval, banks are blocking the password workaround, and in the EU the two-factor requirement adds another brake. Its exemptions aren't suited to outflows in the billions.

Two-factor approval only prevents an autonomous run, though, not a human one. If an agent pings you at 3 a.m. saying "Your bank is wobbling, tap here," you'll approve it in seconds. Instant payments, which every bank in the eurozone now has to offer, then deliver the money in seconds. SVB shows that people with smartphones don't need an agent for that. It's still unclear how the EU will regulate future agent mandates. And verifiable evidence for the bank reports Barr cited is still missing.

Sources

#ai agents#banks#financial stability#psd2#regulation#fed

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.