Deepfake Identity Fraud in U.S. Financial Institutions: FinCEN Red Flags, KYC Controls and SAR Reporting

FinCEN's alert on deepfake media is not new. It was issued on November 13, 2024 as FIN-2024-Alert004, and it was still listed among FinCEN's alerts as of November 13, 2026. What has changed since then is the setting around it. Regulators have clarified how digital credentials fit into bank customer identification, FinCEN has narrowed some SAR expectations, and industry groups have published playbooks on AI-driven identity attacks. A team reading the alert today needs that context as well as the alert itself.

This guide is for compliance and BSA officers at banks, fintechs, money transmitters and crypto businesses. It keeps three things apart: what FinCEN listed as red flags, what FinCEN recommended, and what other credible sources add. It is also explicit about where legal obligations differ by type of institution, because they do.

What FIN-2024-Alert004 is, and what it is not

FinCEN issued the alert after seeing, from 2023 onward, more suspicious activity reports that described suspected deepfake use, especially altered or fabricated identity documents used to get past verification. It defines deepfake media as AI-generated content that looks real but is not authentic, including images, video, audio and text. The alert itself and the press release are both dated November 13, 2024.

An alert is guidance. It describes typologies, lists red flags and restates existing Bank Secrecy Act reporting duties. It does not create a new regulation, and nothing in it requires an institution to buy detection software. The press release lists casinos, depository institutions, insurance, money services businesses, mortgage companies, precious metals dealers and securities and futures firms among the industries it speaks to. That breadth matters later, because the legal tools each sector has are not the same.

We found no rescission or replacement of the alert. FinCEN's SAR key-term page, last updated in October 2026, still lists it.

What has happened since

None of these items converts the 2024 alert into a mandate. Together they show where examiners' attention is likely to be: how an institution proves a customer is who they claim, and whether its reporting reflects what it actually saw.

How deepfakes reach the customer lifecycle

FinCEN frames identity exploitation in three places: account opening, account access, and during transactions. Deepfakes touch all three, through different techniques.

• Altered or generated identity documents. The alert describes edited or wholly synthetic images used on driver's licenses and passport cards and books.
• Synthetic identities. A generated face is paired with stolen or fabricated personal data to create a customer who never existed. The alert links accounts opened this way to check, card, loan, unemployment and authorized push payment fraud proceeds, and to funnel accounts that collect deposits from many places.
• Live impersonation. Video checks can be defeated with tools that feed pre-recorded or generated video into the verification session in place of a camera.
• Cloned voices. Used against customers (family-emergency scams, aimed often at older adults) and against staff (an executive's voice instructing a large transfer).
• Generated text. Profile answers and chat responses written by a model, which the alert treats as a signal in its own right.

The red flags, and where each one comes from

FinCEN is explicit that no single red flag is necessarily indicative of illicit activity and that institutions should weigh all the facts and circumstances. The alert's bulleted list of nine red flags can be read in four groups; the content below follows the alert's text.

What the alert recommends, as distinct from its red flags. Re-review account-opening documents when deepfake use is suspected; use reverse image searches, metadata analysis and detection tools; apply enhanced due diligence to accounts with other suspicious indicators; require multifactor authentication, including phishing-resistant options, and live audio or video verification; watch for evasion behaviour and virtual camera tools; manage identity verification vendors under sound third-party risk practice; and track customer and employee exposure to AI-enabled social engineering.

What the alert does not contain. Several controls that practitioners now treat as standard come from other sources, not from FinCEN. The industry papers produced under Treasury's initiative, for example, describe attacks on the camera feed itself and the use of cryptographically bound credentials such as passkeys and mobile driver's licenses, on the reasoning that a generated face cannot prove possession of a private key. Those papers are voluntary guidance from industry bodies. Treat them as design input, not as examination criteria.

Identity verification and onboarding controls

Start by separating three questions that a single “verified” result tends to blur: whether the document is genuine, whether the data on it is accurate, and whether the person presenting it is the person it describes. A deepfake attack usually defeats the third while leaving the first two looking fine. Our guide to spotting altered IDs and residence permits covers the document side in depth, from a Canadian regulatory angle; the same discipline applies in the U.S.

A workable onboarding design layers independent checks, so that defeating one does not defeat all of them:

• Document authentication that checks security features and machine-readable data, not only text extraction.
• Face match and liveness, with a question to put to every vendor: does it detect video injected into the session, or only a spoof held up to the camera?
• Corroboration from outside the session, such as authoritative data sources, which does not depend on anything the applicant can generate.
• Device, network and location signals compared with the identity claimed.
• Authentication enrolment at onboarding, preferring phishing-resistant methods, so the credential that follows the identity is stronger than a password and a text message.
• Step-up review by a trained person where the layers disagree.

Design the fallback as carefully as the main path. FinCEN points to attempts to change communication method during a live check as a warning sign. That only works as a control if your own process does not offer an easier route. If a failed video check can be “resolved” by a call or an emailed photo, the weakest channel becomes the real control. Any fallback should be at least as strong as the path it replaces, and its use should be logged and reviewed.

Treat detector output as a signal, not a verdict. FinCEN lists detection software among the tools and notes that a flag is a red flag. It does not state how accurate any tool is, and this article makes no claim about accuracy either. Measure it yourself: run known-genuine and simulated-attack samples through the vendor's system, look at false rejections as well as false acceptances, and repeat when the vendor updates its models. Genuine customers will trip flags: older phones, screen-reader and streaming software, camera filters and poor lighting all do. Provide a manual review route so that a flag does not become an unexplained denial.

Digital credentials. Since September 8, 2026, banks and credit unions may treat an unexpired government-issued mobile driver's license or similar verifiable digital credential as documentary identification under the CIP rule, if their written program and technology support it. It is not mandatory, and indicators of fraud in a credential must still be considered in forming the reasonable belief the rule requires. Check that your written CIP describes digital formats before relying on them. A regulator elsewhere has gone further on liveness: Switzerland's FINMA revised its video and online identification rules effective November 1, 2026, which we cover in our FINMA article. It is not U.S. law, but it shows where a major supervisor has landed.

Vendors. FinCEN's alert points institutions to third-party risk management practice and cites the 2023 interagency guidance on third-party relationships. Under the bank CIP rule, reliance on another institution is limited to one that is itself subject to an AML program rule and regulated by a federal functional regulator, with an annual certification. In our reading, an identity-verification vendor is normally a service provider, not that kind of institution, so the bank or fintech remains responsible for the outcome. Ask for test evidence, change notifications and the right to see how the model performs on your own population.

Monitoring after the account is open

Onboarding controls will miss some synthetic customers, so monitoring has to assume that. Three design choices help.

Carry the onboarding evidence forward. An account that cleared verification only after a manual override or a fallback route should enter monitoring with a higher risk setting than one that passed cleanly. Most systems do not do this by default, because identity verification and monitoring sit in different teams and tools.

Look across accounts, not only within them. The alert notes coordinated activity across multiple similar accounts. Link analysis on device identifiers, network addresses, payout destinations and contact details finds clusters that account-by-account rules never will, because each account looks modest on its own.

Treat the early life of an account as its own risk window. The alert's activity indicators concentrate on new or lightly used accounts. Setting a distinct early-life scenario set (and a defined review period of your own choosing, since FinCEN specifies none) lets you tighten velocity and payee-risk thresholds when behaviour is least established. For a general treatment of rule design, alert investigation and tuning, see our guide to AML transaction monitoring.

Two cautions. FinCEN's October 2025 FAQs say a SAR is not required on activity near the currency transaction reporting threshold unless there is information that it is designed to evade reporting, so the funnel-account pattern in the alert should be assessed on its facts, not filed by reflex. And payment instructions that arrive by voice or video should be confirmed out of band, using contact details already on file and never those supplied in the request, with dual approval above a threshold you set.

Escalation: suspected manipulation versus confirmed fraud

The hardest practical question is how much evidence is enough to act. A useful way to answer it is to name the rungs and attach a default response to each.

Two points sit behind the table. First, “suspected AI manipulation” is not the same as intent to defraud: a genuine applicant using a filter or an enhancement app can produce the same artefacts. Second, the person who decides should not be the person under pressure to close the queue. Define who can approve an override, how fast each rung must be cleared, and what the reviewer must record. Where an escalation path crosses teams, our overview of the three lines of defense helps to place ownership.

SAR reporting for deepfake-enabled activity

The key term. FinCEN asks institutions to include FIN-2024-DEEPFAKEFRAUD in SAR field 2 (“Filing Institution Note to FinCEN”) and in the narrative, and to add any key terms for the underlying typology in the narrative as well. The word is “requests”: this is a filing instruction attached to an alert, not a regulation. FinCEN's key-term page lists the term and date but gives no field instruction, so the alert is the source for where it goes. Copy the term exactly as printed, hyphens included. At least one republished summary drops a hyphen, and a mistyped key term will not match whatever FinCEN searches for.

Whether a SAR is required is a separate question from the key term. The key term only labels a report you were already going to file. The obligation comes from your own SAR rule:

• Banks file when a transaction involves or aggregates at least $5,000 and the bank knows, suspects or has reason to suspect one of the listed conditions. The SAR is due within 30 calendar days of first detecting facts that may justify filing, extended by up to 30 further days if no suspect has been identified, and never later than 60 days after initial detection (31 CFR 1020.320).
• Money services businesses file at $2,000, covering transactions conducted or attempted, within 30 calendar days. Money order and traveler's check issuers use $5,000 when the activity is found in clearance records. MSBs may also file voluntarily where no filing is required (31 CFR 1022.320).
• Both rules call for keeping a copy and supporting documentation for five years, and both keep the report confidential.

If you decline an application before any funds move, whether it reaches a threshold or another reporting condition is a judgement for your BSA officer. Make that judgement deliberately and write it down. FinCEN's October 2025 FAQs state there is no explicit legal requirement to document a decision not to file, but a short rationale is what bank partners and examiners usually ask to see.

What a useful narrative contains. Describe how the identity check was defeated or appeared to be, in observable terms: the session logs, the device and network identifiers, the document numbers as presented, the accounts that share them, and where the money went. Name any tool output as an observation (“the vendor's check returned a high manipulation score”), not as proof. Use the word “suspected” where manipulation has not been confirmed. Say what you did next. A narrative that says only “deepfake detected” gives investigators nothing to follow.

Filing in good faith is protected: the MSB rule cites the safe harbor in 31 U.S.C. 5318(g)(3), and the bank rule is built on the same statute. Do not wait for certainty about the technology before reporting activity that otherwise meets your standard.

Legal obligations by institution type

The alert's red flags are useful to every institution. The legal source of the duty, and the standard you will be examined against, is not the same for all.

The consequence is practical. A bank can be examined on whether its verification supports a reasonable belief of true identity under a rule that names documents and non-documentary methods. An MSB is judged on whether its risk-based program verifies customers adequately for its business, and the independent review has to test that. Neither can borrow the other's rule as a defence.

For MSBs and fintechs writing or updating program documents, our BSA/AML compliance program services page describes how customer identification, SAR procedures and risk assessment fit together. For businesses that handle digital assets, where withdrawals leave quickly and cannot be recalled, our crypto AML compliance services cover onboarding, due diligence and monitoring by business model. To check which crypto obligations are settled law and which were proposals, see our note on FinCEN's October 2026 withdrawals.

Three illustrative scenarios

The following are constructed for illustration. They are not real cases and are not drawn from any enforcement action.

Scenario 1: the ambiguous selfie. A wallet applicant's face match scores low, and the vendor flags possible manipulation. The applicant used a beautifying filter on an older phone. A rule that auto-declines would reject a genuine customer; a rule that auto-approves would pass a fake. The reviewer asks for a re-capture on a different device, corroborates against an independent data source, and approves with the override logged. No SAR follows. The record shows the flag, the review and the reason.

Scenario 2: the cluster. A money transmitter sees eleven new accounts in two days under different names. Each submitted a clean-looking document, and each is individually below the amounts that draw attention. Link analysis shows one device fingerprint behind several of them and a common payout destination. The team freezes the cluster, preserves the logs and files one SAR that describes the group, with the key term in field 2 and the narrative. The case depended on joining the accounts together; no single account would have raised an alert.

Scenario 3: the urgent call. A caller who sounds like the CEO asks a crypto company's finance team to move a large stablecoin balance to a new address. The employee follows the procedure: call back on the number held on file, then dual approval. The CEO knows nothing about it. Nothing leaves the building. The firm records the attempt, tests whether it meets its SAR criteria, and uses it in staff training. The control that worked was a procedure, not a detection model.

Governance, vendor oversight and testing

Three artefacts show an examiner or bank partner that deepfake risk is managed, not just known about.

• The risk assessment treats identity-proofing channels as a risk factor, and is revisited when a channel, a vendor or a verification method changes.
• The written procedures say who enters the SAR key term, who approves overrides, and what the verification team does when an automated flag and a human judgement disagree.
• The test record shows that someone independent has checked whether these controls operate. For MSBs, 31 CFR 1022.210(d)(4) requires an independent review scaled to risk, performed by someone other than the designated compliance officer. A sample of onboarding decisions that includes overrides and fallback approvals will tell you more than a review of the policy text. Our BSA/AML audit checklist shows how customer identification and SAR controls are tested, and our independent BSA/AML audit service delivers the review.

It is also worth knowing what is still unsettled. FinCEN's April 2026 program proposal would change how program effectiveness is judged, but it is a proposal; the programs now in force are those described above.

Where to start

1. List every point at which a person proves who they are, including call-centre recovery and manual overrides, and note which ones depend on a camera or a voice.
2. Pull the last 90 days of verification overrides and fallback approvals and read a sample.
3. Confirm that someone owns key-term entry on SARs and that the term is spelled correctly in your template.
4. Test your vendor on your own data instead of relying on its brochure.
5. Read your written CIP or program language and check that it covers digital credentials, if you intend to accept them.

If you run an MSB, fintech or crypto business and want an outside view of how these controls would hold up, the pages linked above describe how ComplyFactor works with U.S. businesses.

Frequently Asked Questions

How long do we have to keep verification evidence, and does that include selfies and video?

The bank CIP rule requires a record of the identifying information obtained, a description of the documents relied on, the methods and results of any non-documentary verification, and how discrepancies were resolved. Identifying information is kept for five years after the account closes. The other records are kept for five years after they are made. The rule does not say that a selfie or session video itself must be kept. Whether to retain biometric images is a policy decision, and one that biometric-privacy and data-protection laws, which vary by jurisdiction, may restrict. Settle retention with counsel before you build the workflow, and make sure a deletion schedule cannot erase material you need for a SAR file, which has its own five-year retention.

Does the SAR key term apply only to fake-ID account openings?

No. FIN-2024-Alert004 covers two families of activity: deepfake identity documents used to defeat verification, and deepfake audio and video used in phishing, business email compromise and impersonation scams. Its SAR request is worded around activity connected to the alert, not around the onboarding stage, so a customer who is the victim of a cloned-voice scam can fall within it. If the activity also fits another FinCEN alert, include that alert's key term in the narrative as well.

If a flagged account keeps operating after we file, do we have to keep refiling?

Not in the way many teams assume. FinCEN's October 2025 FAQs say an institution need not run a separate review of a customer or account after filing a SAR to see whether the activity continues; it may rely on risk-based policies and controls designed to detect and report suspicious activity. Institutions that prefer the older practice (a continuing-activity SAR about every 90 days) may still follow it. Whichever you choose, write it into your procedures so reviewers apply it consistently.

Can we share what we learn with other institutions?

Voluntary information sharing under Section 314(b) of the USA PATRIOT Act is the usual route, and FinCEN's September 2026 scam-centre alert encourages it. Whether a particular deepfake case falls within the scope of 314(b) depends on its facts and on whether your institution is registered to share, so take advice on a case-by-case basis. Sharing underlying facts and identifiers with a registered peer institution is different from disclosing a SAR or the fact that one was filed. The confidentiality rules prohibit that disclosure to the customer and also to other 314(b) participants.

What evidence should we preserve when we suspect an attack?

Preserve it before you close the account or let routine deletion run. That means the original uploaded files with their metadata, the raw verification session record, the vendor's full response, device and network logs with timestamps, and the reviewer's notes. Record who handled each item and when, and hold copies unaltered. If a case may be referred to law enforcement or a bank partner, evidence you can show was not changed is more useful than a screenshot.

Official references

FinCEN and Treasury

• FIN-2024-Alert004: FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (Nov. 13, 2024)
• FinCEN press release, Nov. 13, 2024
• FinCEN SAR advisory key terms (last updated October 2026)
• FinCEN and banking agencies, Frequently Asked Questions Regarding Suspicious Activity Reporting Requirements (Oct. 9, 2025)
• FinCEN news release, Sept. 8, 2026: FAQs on verifiable digital credentials under the CIP rule
• FIN-2026-Alert004, Federal Student Aid fraud (July 24, 2026)
• FinCEN, FIN-2019-G001: Application of FinCEN's Regulations to Certain Business Models Involving Convertible Virtual Currencies (May 9, 2019)
• Federal Register, 91 FR 18704: AML/CFT Programs notice of proposed rulemaking (Apr. 10, 2026)
• U.S. Department of the Treasury press release, Feb. 18, 2026, on the conclusion of its AI initiative

Regulations

• 31 CFR 1020.220: bank CIP
• 31 CFR 1020.320: bank SAR
• 31 CFR 1022.210: MSB AML program
• 31 CFR 1022.320: MSB SAR
• 31 CFR 1010.230: beneficial ownership of legal entity customers
• 31 CFR 1023.220: broker-dealer CIP