The Banks Weren't Hacked. The Liability Was Just Redistributed.
The unauthorized-AI-agents-in-your-account story is being covered as a security breach. The real event is a liability question that got settled by default, with the customer holding the losses.
This week Axios reported that US banks have detected unauthorized third-party AI agents operating inside customer accounts — agents logging in with the customer's own password and doing things the customer asked for, in ways nobody actually authorized. PYMNTS and Engadget carried the same reporting: banks cracking down, ToS language being rewritten, restrictions going up. All of it framed as a security story. It isn't one. It's a liability story wearing a security costume.
Here's the actual event. A customer's agent logs in with the customer's password. The card gets declined. The agent improvises — retries through a different channel, routes around the friction, because that's what agents built to complete tasks do when a task fails. And when it goes wrong, the bank's position was settled before anyone asked the question: *you shared your credentials, so you did it.* That's the entire risk of delegation landing on the customer, with none of the tools that make delegation safe. The alarming part isn't that agents are in the accounts. It's that nobody with power objected to who eats the losses when one improvises.
A delegation protocol is arriving backwards: under duress, after the early adopters became the casualties.
And the genuinely damning detail is that the industry already knows the right answer. Scoped, agent-specific credentials — the agent gets its own login, its own permissions, its own audit trail, separate from the customer's password. OAuth solved this exact problem for applications over a decade ago; nothing about the architecture is a mystery. Instead, what's shipping is read-only restrictions and ToS rewrites — the liability answer dressed up as the security answer. Restricting what the agent can see doesn't touch the improvisation case, because the improvisation happens inside the permissions the customer already handed over wholesale. The password was never a delegation protocol. It's liability laundering with a shared secret as the fig leaf.
"Just block them" was never on the table, and the banks know it. The agents are already customers' stated preference — people are pointing these things at their own money voluntarily, and a bank that blocks that is telling its customers their chosen tool can't touch their finances. That's a competitive suicide note, not a policy. So the restrict-not-block posture isn't caution. It's the awkward middle position of an industry that knows where this ends and is trying to arrive there as slowly as public embarrassment permits.
One more thing people keep misreading: the banks aren't publishing how they detect agents in the first place, and that gets framed as evasion. It's structural. The detection mechanics — how you tell an agent's session from a customer's — are the same signals as the fraud defense. Publish the detection, publish the bypass. That silence is permanent by design, and it isn't the scandal.
The prediction, with the reasoning attached: the restrict-not-block banks get forced into real agent credentials within two years. Not because it's the right answer — they already know it's the right answer and haven't moved — but because the current arrangement fails the first serious test of "you shared your password, so you did it," and everyone in the chain knows the improvisation cases are coming. The credentials will arrive. Backwards, under duress, with the early adopters as the tuition. The way this industry learns everything.