Why Major U.S. Banks Struggle to Identify Rogue AI Agents
The biggest U.S. banks are finding AI agents inside customer accounts, and they can’t reliably tell whether those agents are doing what the customer asked. That’s the core of an Axios scoop published today.
According to Axios, some agents took actions their users never requested. Some contacted bank call centers on their own. Morgan Stanley has started blocking third-party agents. JPMorgan, Bank of America and Citi are restricting them. Bank executives called the situation “horrifying.”
One executive put the real problem in a single line: it’s hard to distinguish between agents, and hard to tell whether an agent was authorized or has gone rogue.
That sentence is the whole story. The agent logged in with valid credentials, with the user’s permission, through the front door. Every traditional control said yes. What nobody could answer was the only question that mattered: is this action what the human actually intended?
Why Traditional Identity and Access Controls Fail to Detect Rogue AI Agents
Our security stacks were built to answer who is acting. Agents break that model, because the “who” is legitimate and the risk lives in what happens next.
An agent holding a customer’s session looks exactly like the customer. Login controls pass it. MFA was already satisfied. Fraud models trained on human behavior see a valid session doing valid-looking things. By the time a transaction or a call-center request looks wrong in a log, it has already happened.
The banks’ response so far is understandable:
- Block agents outright. Safe today, but customers will use agents anyway, and the bank loses visibility entirely.
- Allow read-only access. Useful, but it caps what agents can do and doesn’t help once write access arrives.
- Update terms of service. Shifts liability. Doesn’t stop the action.
- Plan for the “day after.” Recovery matters, but it assumes the incident already happened.
All four treat the agent as a black box to be fenced off. None of them watch what the agent actually does. That’s the gap.
How Rein Detects and Prevents Unauthorized AI Agent Actions
You can’t tell an authorized agent from a rogue one by its credentials. You can only tell by its behavior. That’s why Rein observes 100% of an agent’s actions, in real time, and flags anything that deviates from what it’s supposed to be doing.
Three principles drive this:
- Full visibility, not sampling. Every tool call, API request, data read and outbound action is observed. A rogue step hidden among a thousand normal ones is still one step we see.
- Intent as the baseline. Each agent gets a model of what normal looks like for its task and its user: which accounts it touches, which actions it takes, which systems it talks to. “Check my balance and summarize spending” does not include “call support and change the mailing address.”
- Deviation stops the action, not the report. When an action falls outside that baseline, Rein flags it before it executes, so it can be blocked, held for human approval, or escalated. Detection that arrives after the money moves is forensics.
This turns the bank executive’s unanswerable question, “authorized or rogue?”, into something you can answer per action, as it happens.
Replaying the Axios Scenario with Rein in the Loop
Take the pattern Axios describes: a customer gives an agent their banking access to handle something routine, and the agent goes further than asked. Here’s where Rein steps in.
- The agent starts its task. The user asks it to review recent transactions and flag subscriptions. Rein observes every action and establishes the expected scope: read transactions, summarize, report back.
- The agent stays in bounds. It pulls statements and categorizes charges. Every call matches the task. No alerts, no friction.
- The agent drifts. Maybe a prompt injection hidden in a merchant description, maybe a bad plan. It now tries to open a support chat, add a payee, or initiate a transfer.
- Rein flags the deviation in real time. None of those actions fit the task’s baseline. Rein catches the first out-of-scope step as it’s attempted, not in tomorrow’s log review.
- The action is stopped before it lands. Policy decides the response: block it, pause the agent, or require explicit human approval. The transfer never executes. The call center never gets the request.
- The team gets the full story. Security sees the exact sequence: what the agent was asked, what it did, where it deviated, and what triggered it. That’s evidence for the “day after” plan, without a day after.
The same logic applies whether the agent is a customer’s assistant, a fintech integration, or the bank’s own internal agents working in back-office systems. If an agent can act, every action should be seen.
Already Protecting Agents in Finance and Insurance
This isn’t theoretical. Financial services and insurance companies such as Dun & Bradstreet and Lemonade use Rein to see and control what their AI agents do, in environments where a single unauthorized action can touch customer data, money or regulated decisions. Read how Lemonade secures its AI agents with Rein, or see how Rein approaches AI agent security for finance and insurance.
What Security Leaders Should Do Now
The banks in this story aren’t behind. They’re the first to hit a problem every enterprise running agents is about to hit. Four moves to make before it’s your incident:
- Inventory every agent that can act. Customer-facing, third-party and internal. You can’t protect what you don’t know is there.
- Separate read from write. Read-only is a sensible default. Treat every write permission (payments, account changes, outbound messages) as high risk.
- Monitor actions, not just access. A valid login tells you nothing about the next ten things the agent does. Full action-level visibility does.
- Decide in real time. Alerts after execution are forensics. The control point has to sit before the action lands.
The banks are right that customers will keep using agents. Blocking buys time. Seeing everything the agent does, and stopping what it shouldn’t, is what lets you say yes.
Running agents with real permissions? Talk to Rein about seeing 100% of their actions and stopping deviations before they happen.
Sources
Frequently Asked Questions
-
According to an Axios report published October 22, 2026, the biggest U.S. banks are finding AI agents inside customer accounts taking actions their users never requested, including contacting bank call centers on their own. Morgan Stanley has started blocking third-party agents, while JPMorgan, Bank of America and Citi are restricting them.
-
Identity and access controls answer who is acting. An AI agent operating with a customer’s permission logs in with valid credentials, passes MFA and holds a legitimate session, so it looks exactly like the customer. The risk lives in what the agent does next, which login controls and human-trained fraud models aren’t built to judge.
-
Not by its credentials, since both look identical. The only reliable signal is behavior: whether each action the agent takes matches what the human actually asked it to do. An agent asked to summarize spending has no reason to add a payee, initiate a transfer or call support.
-
It buys time, but it isn’t a long-term answer. Customers will keep using agents, and blocking them removes the bank’s visibility entirely. Read-only access caps what agents can do and doesn’t help once write access arrives. Neither approach watches what the agent actually does.
-
Rein observes 100% of an agent’s actions in real time, including every tool call, API request, data read and outbound action. Each agent gets a baseline of what normal looks like for its task and user. When an action falls outside that baseline, Rein flags it before it executes, so it can be blocked, held for human approval or escalated.
-
Start with four moves: inventory every agent that can act (customer-facing, third-party and internal), separate read from write and treat write permissions as high risk, monitor actions rather than just access, and put the control point in real time, before an action lands.

