Treasury's March 2024 AI Report and What Came Next

Updated October 2026. First published April 2024.

On March 27, 2024, the U.S. Treasury published Managing Artificial Intelligence-Specific Cybersecurity Risks in the Financial Services Sector. Executive Order 14110 required it within 150 days. Treasury's Office of Cybersecurity and Critical Infrastructure Protection wrote it from 42 in-depth interviews conducted in late 2023 with banks, credit unions, technology firms, data providers, and anti-fraud vendors.

The report imposes no requirements. It describes what firms told Treasury and ends with a list of next steps. Two and a half years later, most of that list has been worked on, so the report reads best alongside what came after it.

What the Report Found

Firms weren't new to AI. Interviewees said they had used machine-learning fraud tools for more than a decade, and AI-based anomaly detection was already built into their endpoint, intrusion-detection, and data-loss tools. Generative AI was the part they were careful with. Most of the firms Treasury interviewed hadn't used public generative AI services and ran enterprise versions inside their own cloud tenant instead. The low-risk use they were putting first was code assistants.

Treasury's read of the threat was plain. Attackers using new AI tools would likely have the advantage at first. Most successful attacks so far had been social engineering and identity spoofing, and firms consistently named AI-generated impersonation as a chief concern. The report cites a 2024 case in Hong Kong where an employee transferred $25 million after a video call with a deepfake chief financial officer. It warns that even a live video call with a known client may no longer be enough to verify identity.

For attacks on AI systems themselves, the report lists four categories drawn from NIST's work on adversarial machine learning:

Data poisoning: corrupting the training data or the model's weights.
Data leakage during inference: pulling confidential data out of a model by querying it.
Evasion: crafting an input that makes the model produce the output the attacker wants.
Model extraction: building a working copy of a model by querying it over and over.

The finding with the most consequence was about size. Large firms had their own data and their own people; one told Treasury it had about 400 employees working on fraud-prevention AI. Smaller firms relied on vendors and didn't have enough fraud history to train a model even if they could build one. The sector has shared cyber threat data through FS-ISAC since 1999, but there was no clearinghouse for fraud data. Fraud blocked at large firms, the report says, would likely move to smaller ones.

The Vendor Questions

For a broker-dealer that buys its AI instead of building it, section 5.5 is the most useful page in the report. Treasury suggests asking AI vendors to:

tell you when they change a product or service that uses AI;
disclose how much of the product uses AI, and notify you of material changes;
describe the model and data lifecycle where AI is significant to the product;
explain how the AI could affect your customers, and how you'd explain that to them;
describe patching and vulnerability assessment for the infrastructure that hosts the AI;
name any third-party AI models built into the product.

The last item matters more than it looks. Many interviewees said their vendors got the AI from other vendors, so the model inside a product may belong to a company you've never reviewed. FINRA's outsourcing guidance in Regulatory Notice 21-29 still applies: outsourcing a function "does not relieve members of their ultimate responsibility" for compliance. We covered FINRA's third-party expectations in The Vendor Compliance Time Bomb.

What Followed

The report's last section named ten open problems. They included a common AI vocabulary, a financial-sector version of the NIST AI Risk Management Framework, "nutrition labels" describing the data behind vendor AI, explainability, the fraud data divide, and digital identity. This is what has happened since:

In June 2024, Treasury asked for public comment on AI in financial services more broadly. Its December 19, 2024 report summarized 103 comment letters and recommended more coordination among regulators.
On January 20, 2025, Executive Order 14148 revoked EO 14110, the order behind the March 2024 report. The report itself was never binding, so nothing in it lapsed.
In February 2026, the Artificial Intelligence Executive Oversight Group, a partnership of the Financial and Banking Information Infrastructure Committee and the Financial Services Sector Coordinating Council, began releasing a series of resources. Treasury announced the first two on February 19: an AI Lexicon and the Financial Services AI Risk Management Framework, which adapts NIST's framework to financial firms. The rest of the series covers AI attacks on identity and authentication, explainability, data nutrition labeling, and AI-enhanced fraud.
From March to May 2026, Treasury held four AI roundtables, its AI Innovation Series, and published readouts in June. No new rule came out of them.

Set the two lists side by side and the 2026 papers track the 2024 next steps almost item for item: lexicon, risk framework, nutrition labels, explainability, fraud, identity. The framing changed in between. The 2024 report answered an order about safe and secure AI. Treasury presents the 2026 materials as support for the White House's July 2025 AI Action Plan and as tools to help firms adopt AI with more confidence. All of it is still guidance.

What a Broker-Dealer Should Take From It

Nothing here creates a new obligation for a broker-dealer. Regulation S-P, FINRA Rule 3110, and the outsourcing guidance apply to an AI tool the way they apply to any other system. What the report offers is a checklist written by firms that had already been through the exercise. Add the vendor questions above to your due diligence. Ask who supplies the model inside each product. And treat a voice or video request to move money or change settlement instructions as unverified until someone confirms it through a channel you already trust.

This post is for informational purposes only and does not constitute legal advice. For guidance on specific regulatory obligations, consult your counsel or compliance advisor.