Table of Contents

Banks Are Blocking the Wrong Agents

Bot detection vendors promise to identify AI agents. Banks trust them. But what if those defenses are missing most of the agents coming through the door?

Banks have decided to block AI agents. There’s just one problem: the technology they’re relying on may be detecting only a small fraction of them.

In a single week this October, four of the largest US banks announced restrictions on third-party AI agents accessing customer accounts. Several major Canadian banks followed. The message was clear: AI agents aren’t welcome inside customer accounts unless banks can control them.

Fair enough. But here’s where the story gets interesting: Banks aren’t building their own agent detection technology. Most rely on established bot detection and fraud prevention vendors, many of whom now advertise AI agent detection capabilities.

The vendors say they can detect agents. Banks deploy their products, configure blocking policies, and receive dashboards showing how many agents have been stopped. Everyone feels protected. Except the dashboards may be showing only a small fraction of the agents actually accessing those banks.

What the dashboard doesn’t show

In October, the CEO of a large financial institution discussed receiving daily reports showing how many AI agents the company had blocked. The numbers were small, leading him to suggest that customers weren’t yet using agents extensively for financial activities.

But there is a fundamental flaw in that reasoning: A security dashboard doesn’t show how many agents visited your website. It shows how many your security technology managed to recognize. Those are two very different numbers.

Imagine a bank receives 10,000 agent-driven sessions, but its bot detection technology identifies only 1,000. The dashboard proudly reports 1,000 detected agents. The bank may block all of them and celebrate a 100% blocking rate. Meanwhile, 9,000 agent-driven sessions go unrecognized.

The numbers are illustrative, but the problem is real: you cannot measure the success of a detection system by counting only what it detects. Yet that is effectively what happens when banks interpret low agent-blocking numbers as evidence of low agent adoption.

The question isn’t how many agents were blocked. It’s how many got through.

A different generation of automation

For years, bot detection vendors have built sophisticated systems to identify automated traffic. Their technologies examine browser characteristics, network signals, interaction patterns, mouse movements, typing behavior, and other indicators that distinguish automated software from human users. These approaches have delivered real value against traditional bots.

But the arrival of AI agents changes the game. Unlike traditional automation, modern AI assistants can operate through ordinary browsers, navigate websites dynamically, and interact with applications in ways that don’t necessarily expose the signatures conventional bot detection was designed to recognize.

Some agents openly identify themselves. Others don’t. Some use recognizable automation frameworks. Others operate through browser environments that are much harder to distinguish from legitimate customer activity. And many agent platforms can evolve their behavior when they encounter detection mechanisms.

The industry is trying to solve a new identification problem using technologies largely developed for a different generation of automation. That doesn’t mean traditional bot detection is useless. It means detecting conventional automation and reliably identifying modern AI agents are not the same problem.

Yet vendors increasingly market both capabilities under the same umbrella. And banks may be assuming that a product capable of stopping traditional bots is equally capable of recognizing AI agents.

That assumption needs to be tested, not trusted.

The easiest agents to block

Here’s the irony: An AI agent that openly declares its identity is relatively easy to recognize and block. An agent that conceals its identity, uses a conventional browser, or operates through an existing customer session can be significantly harder to identify.

That creates a perverse outcome: The agents that cooperate with websites by identifying themselves are often the easiest to stop. The agents that don’t declare themselves may have a better chance of getting through.

In other words, banks could be building a security strategy that disproportionately blocks transparent agents while allowing less transparent ones to operate. And because the undetected agents don’t appear in the dashboards, banks may have no idea how much activity they’re missing.

The absence of detected agents is being confused with the absence of agents.

Testing the detection claims

The banking industry has spent decades demanding rigorous testing of cybersecurity products. Banks run penetration tests. They commission independent assessments. They test fraud models against known attack patterns. They measure false positives and false negatives. Why should AI agent detection be any different?

If a vendor claims to detect AI agents, banks should demand answers to a few straightforward questions:

  • Which specific agent platforms can you identify?
  • Can you recognize agents that don’t declare themselves?
  • Can you distinguish one agent platform from another?
  • What percentage of agent-driven sessions do you actually detect?
  • How does detection perform when an agent operates through an authenticated customer session?
  • Have those detection rates been independently validated?

A vendor showing a dashboard full of blocked agents is not demonstrating comprehensive agent detection. It’s demonstrating that it can detect the agents appearing on that dashboard. Those are not equivalent.

The question banks should ask

When a bank’s security dashboard reports that 500 AI agents were blocked yesterday, the CEO shouldn’t ask whether 500 is a large or small number. The CEO should ask: “What percentage of the agents accessing our bank did we actually detect?”

And if the vendor cannot answer that question with credible testing, the bank has no reliable way to judge whether its blocking strategy works. The biggest problem isn’t that banks are trying to block AI agents. It’s that they may be trusting technology that catches only a fraction of them.

Banks think they’re blocking AI agents. Their vendors may simply be blocking the ones they can recognize. And the agents they miss are the ones that should concern them most.