On 27 October, Bank of America published a rule that sounds obvious and reads, on closer inspection, like a confession. The bank now asks customers, plainly, not to hand their banking credentials to an AI assistant, as TechCrunch reported. Nobody is shocked that a bank guards its passwords. What jars is the timing: autonomous agents already sit inside browsers, and banks are only now drafting rules for software that acts on their customers' behalf.

That's the real story of banking security this week. The perimeter of trust no longer ends at the institution's app. It now includes a third party the bank doesn't control, can't see and never certified.

Banking security: the weak link moved outside the bank

For twenty years, protecting an account meant answering three questions. Is the password strong? Is the connection encrypted? Does the bank spot odd behaviour? All three still matter. None of them is enough.

When an AI agent logs in for you, it handles your credentials in an environment your bank doesn't audit. It can read a balance, start a transfer, change a payee. The bank sees a perfectly legitimate authenticated session. It doesn't see the hand on the keyboard.

That's the ground Decrypt maps out: AI agents are creating a new layer of cybersecurity spending that nobody budgeted for. Firms aren't paying to keep a distant attacker out. They're paying to keep their own automations on a leash. The threat changed address.

What Washington and Ripple are really signalling

The same day, the US Treasury said something that looks unrelated. Scott Bessent framed crypto as a tool for keeping the dollar on top, according to CoinDesk. Reading that as a monetary boast would miss the point. A dollar moving on programmable rails is a dollar whose every move can be traced, signed and verified.

At the same moment, Ripple won final approval for a US trust bank, CryptoSlate reports. The precedent matters more than the licence: a blockchain-native firm has stepped inside the regulated banking perimeter. Compliance is becoming a product, not a cost centre.

The calendar isn't being kind either. A reported flaw in Lightning invoice handling briefly made Bitcoin withdrawals look unpaid, per CryptoSlate. A settlement that appears complete without being complete is every treasury desk's nightmare.

Security is no longer what stops a payment leaving. It is what proves the payment arrived, in the right place, exactly once.

The technical angle: proof over defence

Granted, you can argue these incidents are marginal and that banks have never been better defended. Social-engineering losses are falling in several major markets, and customers of large institutions rarely lose money on confirmed fraud. Fair. It still misses the architecture.

Perimeter defence (firewalls, authentication, anomaly detection) assumes an inside and an outside. AI agents erase that line. An assistant that pays your bills is both your delegate and a potential attack vector. You can't block it without blocking the customer.

Hence the shift toward proof: signed instructions, timestamps, conditional release, tamper-evident logs. A payment is no longer validated by the password that authorised it but by the chain of evidence that travels with it. That's the logic behind escrow-based secure payments: funds move only when a condition is met and can be checked.

A single Lightning invoicing flaw on 27 October was enough to make Bitcoin withdrawals look unpaid, a problem of proof, not of cryptography.

The numbers finance chiefs should sit with

Robinhood reported third-quarter results the same day. Crypto revenue fell 55% year on year, per CoinDesk, even as it beat Street forecasts. Its own chain brought in $26 million, The Block reports.

What does that have to do with security? Something simple. When trading revenue turns volatile, platforms hunt for recurring revenue. The first recurring revenue any fintech has is trust. An account people trust gets kept. An account they doubt gets emptied.

Finance teams understand this, slowly. Here's what the sharper ones have already changed: the assistant never holds the password, it receives a limited, revocable, logged mandate; every payment instruction leaves a verifiable trail for the customer and the bank alike; and a multi-currency account with legible controls outsells an opaque one.

What this means for you

  • Never give banking credentials to an AI assistant, however polite and however neatly built into your browser. Ask for scoped delegation you can revoke in one click.
  • Check that a large transfer leaves evidence you can inspect: reference, timestamp, final status. If your bank can't show it, ask before the next payment, not after.
  • For a remote purchase or a service, prefer a mechanism that releases money only once delivery is confirmed. Belook's Escrow product page shows what that guarantee looks like in practice.

The next test is dated. US midterms and a Federal Reserve decision are looming, with bitcoin hovering near $80,500 according to Bitcoin Magazine. Volatile weeks are when impersonation fraud spikes. So watch one specific thing: the next revision of the Lightning invoicing rules. That's where the technical difference between a payment that looks settled and one that is settled gets decided.

Sources