DARPA’s A3ML program aims to raise the cost of money laundering with privacy-preserving detection, program manager says

DARPA’s anticipatory and adaptive anti‑money‑laundering program, A3ML, is designed to make money laundering more costly and thus less viable, program manager David Rushing Doohurst said in a podcast interview.

"If I ever told you that I would try to eliminate all global moneyaundering, that's goofy," David said, adding that the program instead seeks to "move the supply curve of money laundering to the left" so the price of illicit finance rises and deterrence follows.

A3ML splits work into two technical areas, David explained. One area (TA1) searches for patterns of illicit finance across large, heterogeneous data sets. The other (TA2) plays an adversary role: researchers create hypothetical laundering tactics, techniques and procedures (TTPs) and inject those scenarios into data to test how well TA1 systems detect novel methods. "TA2 metrics are how well can you fool the TA1 performers, but also how realistic is what you've created?" he said.

David emphasized that the program is intentionally built to protect privacy. Rather than centralizing sensitive customer records, A3ML sends analytics to data holders and returns aggregated, abstracted views of suspicious patterns: "Everybody keep your own data. That's fine. We don't want to see your private data," he said. The system reports counts or semantic summaries but not records that would reveal one source's private data to another.

David described A3ML as a research effort focused on technical capability, not policy enforcement. He noted DARPA does not compel downstream agencies or financial institutions to adopt tools: adoption would require later buy‑in from agencies such as the Financial Crimes Enforcement Network (FinCEN) or private‑sector actors.

A3ML is early-stage and uses a recursive red‑teaming design to harden detection against evolving tactics. David called it promising but asked listeners to "check back in a year." The program manager also stressed that even a technically successful DARPA system requires follow‑on programs, agency uptake and corporate willingness to be effective at scale.

The episode suggested A3ML aims to balance improved detection across global datasets with constraints on moving private data, while using adversarial testing to find and close detection gaps. The program’s timeline and eventual operational adoption remain contingent on external agencies and industry partners.