Sumsub published its sixth annual Identity Fraud Report on November 17, and two of its numbers belong side by side. Among U.S. and Canadian businesses it surveyed, 29% verify who authorizes an AI agent, and 9% can suspend one. I think finance firms are working on those two controls in the wrong order. The ability to stop an agent should come first.
The report’s summary draws on a survey Sumsub ran in August 2026 among more than 1,400 fraud and risk professionals and more than 1,500 consumers across 21 markets. Globally, 53% of businesses said they had detected or suspected fraud involving an AI agent. Among U.S. and Canadian businesses the figure was 41%.
The agent in the fraud case is usually a real one
One finding carries my argument. In 49% of the cases where businesses detected or suspected agent fraud, Sumsub says “a legitimate agent had been manipulated by malicious instructions”. In 30%, an agent initiated a purchase, payment or transfer.
Consider what identity verification does in the first group. The agent is the one the customer set up. The human behind it is who they claim to be, and a check on who authorized the agent passes. The fraud arrives later, in the instructions the agent reads. A firm that has spent its budget on proving who stands behind an agent has done nothing about an agent that was honest at onboarding and compromised weeks later.
Suspension works on that case. A firm that can freeze one agent’s access within minutes limits the loss to what the agent did before someone noticed. A firm that cannot is left with the tools it uses for a compromised human account: block the whole customer, or let the agent keep running while the case is investigated.
Why suspension is the cheaper control to build
An agent reaches an account through something the firm already issues. Last week this publication described the four doors in use today, in a feature on how finance firms decide which AI agents can touch an account: a login, a card, an API key or a supplier portal. Each is a credential the firm can revoke. Suspension means giving the agent its own credential, separate from the customer’s, and giving the fraud team a switch for it.
Verifying the authorizer is a larger project. It needs a way to bind an agent to a person, a way to re-check the binding, and agreement with the agent’s developer on how that is presented. Sumsub sells one version of this, a Know Your Agent framework that, by its description, links each agent to a verified human and requires a human check when an agent tries to exceed its permissions. That is a reasonable design. My point is about sequence. A firm with neither control gets more protection per dollar from the switch.
The strongest objection
The best argument against me is that a firm cannot suspend what it cannot identify. If agent traffic arrives through the customer’s own login, the fraud team sees a customer session and has no agent to switch off. On that view identity work has to come first, because it is what makes the agent visible.
I accept half of this. Telling an agent apart from its owner is a precondition, and it is a narrower task than verifying who authorized it. A separate credential per agent gives the firm something to revoke without proving anything about the human. Banks that wrote their agent rules after connecting the rails, a pattern we described in an earlier column, can add that credential without redesigning onboarding.
Networks make the same case
The report’s other headline finding points the same way. Sumsub says the largest single fraud network it detected in any country was in the U.S. and linked 3,173 users, and that only 16% of U.S. and Canadian businesses use fraud network analysis. Erin West, former deputy district attorney and founder of Operation Shamrock, describes in the release why per-account checks miss these networks: “Every account is verified, and every document is real, because the person holding it was recruited, paid, or coerced into opening it, so nothing trips a wire on its own.”
West is talking about people, and the logic transfers to software acting for people. Verification at the door tells a firm who came in. It says little about what that account or agent does afterward. We covered a related problem when FinCEN mapped how scam center proceeds reach US banks in three stages.
What I would do this quarter
A fraud or payments leader can run a short test. Pick one agent that a customer has connected, and time how long the team takes to cut its access while leaving the customer’s own access intact. If the answer is that it cannot be done, the firm sits with the 91% of U.S. and Canadian businesses in Sumsub’s survey that lack a suspend control. I would put the fix ahead of any vendor evaluation for agent identity, and I would write the trigger conditions for using it before the first incident.
Source: Sumsub via PR Newswire
