The swarm that wasn’t
Our engine flagged what looked like a coordinated agent attack on a card program. The account holders had ordered every transaction. What agent-enabled laundering looks like from the authorization path, and why sanctions pressure will bring more of it.
The most instructive case we have worked this year did not start with a breach. It started with a clean onboarding.
The case
We sit in the authorization path of a fintech card issuer. Its onboarding is thorough by any industry standard: document verification, a biometric liveness check, sanctions screening, phone number validation and an IP check. A group of users passed every one of those controls, because they were real people using their own identities.
Once approved, each of them began to create AI agents. The program allows up to ten agents per user. These users created between two and six, and every agent received its own card. The cards were topped up daily with amounts between $20 and $150. From there the agents did two things, every day. They paid each other, and they paid merchants built to look like well-known, well-reviewed businesses, in amounts between $5.99 and $29.99.
Principals: real, fully verified individuals acting in concert · Agents per principal: 2 to 6 (program cap: 10) · One card per agent · Top-ups: $20 to $150 per card, daily · Outflows: agent-to-agent transfers and purchases at lookalike merchants, $5.99 to $29.99 each · Cadence: repeated every day · Ceiling at the top of the range: 6 cards × $150 = $900 per principal per day.
Our engine caught the behavior. Step-up requests, where the engine requires the principal to authenticate a transaction before it proceeds, started rising across the group. Eva, our AML triage agent, opened a case, and the shape of the activity pointed it to a clear first hypothesis: a swarm attack. Many agents, correlated timing, repeated small payments into a narrow set of counterparties. It looked as if someone had taken control of these accounts and was draining them.
The hypothesis was wrong. Every step-up was completed by the legitimate account holder. The principals were not the victims of the swarm. They were running it.
Why the first read was wrong
The first read was reasonable, and the reason it failed is the most useful thing in this case.
Fraud controls and AML controls ask different questions. A fraud control protects the account holder and asks whether the owner authorized the transaction. A step-up is the strongest form of that question, and here it returned the strongest possible answer: the principal authenticated, on their own device, again and again. In a fraud model, a completed step-up clears the transaction.
In an AML model, the account holder is the subject under examination. When the principal is the launderer, every control built to confirm consent will confirm it, because consent is the one thing the launderer has in abundance.
| Control | Result | Why it did not stop the scheme |
|---|---|---|
| Document verification | Pass | Real documents, real people |
| Biometric liveness | Pass | The principals were present |
| Sanctions screening | Pass | No listed parties |
| Phone and IP validation | Pass | Their own phones and networks |
| Agent cap and payment limits | Pass | 2 to 6 agents against a cap of 10; every payment small |
| Step-up authentication | Pass | The principal approved, every time |
Each of these controls worked as designed. Each one examined a single identity, a single agent or a single transaction, and each of those, taken alone, was clean. The scheme existed only at the level above them: a group of principals, their agents and a set of merchants, moving money in the same rhythm.
In classic AML terms the structure is familiar. Placement happens through the daily top-ups. Layering happens through agent-to-agent transfers that mix funds across cards and principals. The purchases at lookalike merchants are where value leaves the program, in amounts and descriptors that resemble ordinary consumer spending. None of it is new. What is new is the operator. A ring used to need discipline to hold a pattern like this steady across dozens of people for weeks. Agents hold the cadence without effort, and the principals only need to approve.
In our research register, this case sits across two money-laundering typologies we had rated anticipated: micropayment mule networks and layering through agent-to-agent chains. We now treat both as observed.
What the identity model changes
We do not score agents as free-standing actors. Every agent is tied to the principal that deployed it, starts from that principal’s standing, and is then measured on its own behavior. The comparison runs both ways: the principal’s score informs each agent, and the agents’ behavior flows back into the principal. That design is what let a swarm alert become an AML case. Looking at principals and agents together makes three things visible that per-card monitoring misses.
Consent without purpose. A principal who approves every step-up while their agents show no plausible consumer purpose is telling you something. In this pattern, a high step-up completion rate across a correlated group is an AML indicator, and the fraud logic that treats it as clearance has to be overridden.
Structure across principals. Independent customers do not top up on the same cadence, keep their agent counts in the same narrow band under the cap, and pay the same lookalike merchants in the same price bands. When they do, the group is the unit of risk, and the review has to be opened on the group.
Circular value. Money loaded onto the cards moved between agents belonging to the same group before leaving through a small set of merchants. Attributing every agent-to-agent hop back to a principal turns many small, unremarkable purchases into one closed loop.
Eva reclassified the case from suspected account takeover to suspected laundering and assembled the record the issuer’s compliance team needs: every authorization, every step-up, every link between agent and principal, signed and hash-chained. Whether to file is the institution’s decision, and it should be. Our job is to make sure the people who make it are looking at the right hypothesis.
Easy to launch, expensive to keep clean
Last week a platform that sells financial-product tooling posted: “Neobanks are great businesses that are easy to run. Today, you can create your own in 15 minutes.”
I spent years building and growing banking-as-a-service, and I want more companies offering financial products, not fewer. Those years also taught me what that sentence leaves out. A card program is fast to launch and expensive to keep clean. Compliance is a cost that grows with the business: monitoring, screening, investigations, audits, the analysts who work the alerts, and the constant investment needed to be ready for the next wave of attempts before it arrives. Every program that grows faster than its controls becomes a candidate target, and launderers look for exactly that profile: new programs, growing quickly, with onboarding tuned for conversion and monitoring tuned for last year’s patterns.
The program in our case did its onboarding right, and the scheme still passed through it. Many programs launched in fifteen minutes will start with less.
Why the pressure is rising
Two forces are pushing more of this activity toward agent payments, one from outside our industry and one from inside it.
The outside force is sanctions enforcement. On September 28, Senator Richard Blumenthal, the top Democrat on the Senate Permanent Subcommittee on Investigations, released a report, first reported by The Wall Street Journal, finding that 84% of 846 Iran-linked sanctioned wallets transacted in Tether’s USDT. The report calls the token “a significant financial lifeline” within Iran’s shadow banking network. Tether disputes the characterization, and about $550 million in Iran-linked USDT has been frozen this year at the request of US authorities. Washington has been closing channels quickly: OFAC designated four Iranian exchanges on June 2, Treasury launched Operation Economic Outcast on August 24, OFAC issued a sectoral determination the next day that reaches anyone operating in Iran’s crypto sector, and on September 17 it added the exchange BitBank.
846 Iran-linked sanctioned wallets reviewed by Senate investigators · 84% transacted in USDT · ~$550M in Iran-linked USDT frozen in 2026 · $104B received on-chain by sanctioned entities in 2025, up 694% (Chainalysis) · 84% of all illicit crypto volume in 2025 moved in stablecoins (Chainalysis).
Nothing in our case links it to Iran or to any sanctioned party. The connection is mechanical. Every enforcement action above works on a known name or a known address, and each one raises the price of the old channels. Illicit money responds the way it always has, by moving to wherever identities are cheap, credible and fast to use. Our case shows what that looks like on a card rail: real, verified people whose identities are multiplied through agents.
The inside force is the models themselves, and it is the reason we started Mandate. The most capable frontier models, and increasingly the open-weight ones, lower the cost of running a scheme and raise its speed. A traditional ring adapts at the pace of people: someone notices a decline, someone changes the amounts, someone recruits another account. An agent-driven scheme can vary its amounts, merchants and timing as fast as the controls push back. Machines adapt faster than the methods our industry built to catch people, so the controls have to sit where the machine acts, in the authorization path.
The crypto lesson
For years, many institutions decided the safest crypto policy was distance. They did not offer it, so they did not build typologies, policies or risk signals for it. Their customers bought and sold crypto anyway, through cards and transfers those institutions processed, and the institutions that had built nothing internally could not see the risk moving through their own rails. Choosing not to participate turned out to be different from choosing not to be exposed.
Agent payments present the same predicament. The debate is usually framed as whether a financial institution should enter agentic payments to capture a new revenue stream. That is the smaller question. Agents are already transacting on cards issued by institutions that never decided to support them. An institution without an agent framework, an agent risk policy and agent-level signals is already in agentic payments, only without instruments.
Regulators will follow
We expect regulators to begin adopting frameworks specific to agent-initiated payments over the coming months and years. Crypto went through the same arc. Bitcoin launched in 2009, and FATF extended its standards to virtual assets in October 2018 and June 2019. For most of that decade, the institutions that had built their own frameworks were the ones prepared when the rules arrived. Agents are being adopted faster than crypto was, and I expect the gap between adoption and regulation to be shorter.
Our reading is that those frameworks will ask for at least three things: that every agent be attributable to a verified principal; that monitoring cover agents and the groups of principals behind them, and not only single accounts; and that every agent decision leave a record an examiner can audit. The purpose is the one that has justified every AML rule before it, which is keeping the financial system from being used against itself.
Every wave builds its control layer
After 2001, sanctions and watchlist screening became a standing function inside every bank, supplied by specialist vendors. When crypto exchanges had to screen addresses, blockchain analytics became an industry of its own. In both cases the new layer was built by specialists working beside the rail.
Agent payments need their own version: a control that ties every agent to a verified principal, watches the group as well as the individual, and treats a principal’s consent as evidence to examine. The institutions that build or buy that capability now will be ready when the frameworks arrive. Our job is to put it in the authorization path before the launderers find the programs that do not have it.
What we build
Mandate Labs is the authorization and fraud-intelligence layer for agent payments. We tie every agent to a verified principal, screen every decision in the authorization path across card and stablecoin rails, and leave a signed record a compliance team can take to its regulator.
Book a demoSources
- The Wall Street Journal, “Senate Investigation Finds Rampant Use of Tether’s Stablecoin by Iranian Regime,” Sep 2026. wsj.com
- Reuters, “Tether USDT aids Iran funding, Senate report says,” Sep 28 2026. via investing.com
- CoinDesk, “Tether is a ‘lifeline’ for Iranian regime, Senate Dems say in new report,” Sep 28 2026. coindesk.com
- Office of Sen. Richard Blumenthal, “Blumenthal Probes Tether On Its Role in Iranian Shadow Banking,” Jun 4 2026. blumenthal.senate.gov
- US Department of the Treasury, “Operation Economic Outcast Disrupts Digital Asset Exchange Enabling the Iranian Regime,” Sep 17 2026. treasury.gov
- Decrypt, “US Can Now Sanction Anyone Operating in Iran’s Crypto Sector,” Aug 25 2026. decrypt.co
- Chainalysis, “2026 Crypto Crime Report: Introduction,” 2026. chainalysis.com
- Chainalysis, “Crypto Sanctions: 2026 Crypto Crime Report,” 2026. chainalysis.com
- FATF, “Guidance for a Risk-Based Approach to Virtual Assets and Virtual Asset Service Providers,” Jun 21 2019. fatf-gafi.org