August 25, 2026Deepfake bank fraud is moving faster than the regulations meant to contain it. Recently, new rules under Article 50 of the EU AI Act came into force. Synthetic content circulating in Europe now needs a machine-readable tag. It’s a sensible step, and as Cyberette founder and CEO Julia Jakimenko wrote in American Banker, it should meaningfully reduce harm from synthetic content used to deceive the public.
For a US bank, it would be easy to read that as a European story. Jakimenko’s own warning is the opposite: American banks would be mistaken to assume the law has nothing to do with them. The Act reaches non-EU providers whose systems touch European users, plenty of US vendors are already in scope, and US regulators are watching the same rollout everyone else is.
Why deepfake bank fraud slips past the rule built to catch it
Article 50 works by requiring disclosure. Synthetic content gets labeled. That’s the right mechanism for deepfakes designed to be seen, shared, and believed by an audience, the kind that spreads misinformation or damages a reputation in public.
Fraud runs on the opposite logic. It needs to go unnoticed. Nobody attempting to defeat a bank’s voice authentication system is going to tag their synthetic audio before feeding it into a call center IVR. That’s not a flaw in the disclosure model, it’s a boundary. The real problem sits on the other side of it, and the US has even less covering that side than Europe does.
A distinction Europe made explicit. The US hasn’t made it at all.
Annex III of the AI Act lists high-risk applications, and biometrics are on that list, with a carve-out. Biometric identification, matching one face or voice against a database of many, is high-risk and regulated. Biometric verification, matching one person against their own stored record, is not.
That second category is what most banks call authentication. It’s the voice match on an inbound fraud call. It’s the selfie check during remote onboarding. It’s the system a customer touches every time they prove who they are to move money.
Europe at least drew the line and named the gap, even if the high-risk obligations covering identification won’t land until 2027 under the Digital Omnibus delay. The US hasn’t drawn an equivalent line yet. There is no Annex III on the table, no proposed high-risk classification for biometric identification, and no path even scheduled that would eventually require verification vendors to demonstrate resilience against manipulation. FinCEN has issued alerts on deepfake-enabled identity fraud, with useful red flags for banks to watch for. But red flags are guidance for spotting an attack after it starts. They ask nothing of the vendors building the verification systems the attack is aimed at.
Jakimenko’s two supporting cases make the exposure concrete, and neither one needed EU jurisdiction to happen. A man in the Netherlands opened dozens of bank accounts by blending his own face with stolen passport photos, exploiting exactly the verification category no regulation covers. Years earlier, a journalist broke into his own US bank account using a free voice-cloning tool, feeding it a handful of samples of his own voice until the system let him in. Both attacks worked the same way: not by defeating security, but by walking through a category regulation had never gotten around to defining.
Why US banks can’t wait for Washington to catch up to Brussels
The EU timeline is the closest thing the industry has to a preview of where US expectations eventually land, and even that timeline runs to 2027 before identification providers face real accountability. Verification, the category actually doing the authenticating on every voice call and every remote onboarding session, isn’t even scheduled for a future compliance deadline on either side of the Atlantic yet.
That’s not a reason to wait. It’s a reason for US banks to hold their own verification stack to a higher standard than any regulator is currently requiring, because when a rule does eventually arrive, it will be built around exactly the failure modes already being exploited today.
This is the argument ValidSoft has been making before any regulator caught up to it. A voice, on its own, is no longer sufficient evidence of identity. Generative tools have made that assumption obsolete. Two questions must be answered before any transaction proceeds: “is this a real human?”, and “is it the right human?”. ValidSoft technology detects synthetic, cloned, and replayed speech in real time. It also matches the verified voice against the right person, passively or actively, against watchlists where needed. Together they close the gap Jakimenko is describing, resilience against manipulation, built into verification itself rather than left to a labeling regime that fraud was never going to obey and a US regulatory framework that doesn’t exist yet.
But closing the identity gap still isn’t the finish line
Even a bank that solves detection and identity perfectly is still exposed. A real, correctly identified human can be coerced. An authorized push payment can be approved by the right person under pressure from a scammer on another line. An AI agent acting with delegated authority can execute outside its actual mandate while every identity check along the way returns a clean result.
Authentication is not authorization. Identity is not consent. Presence is not mandate.
This is where ValidSoft’s VoiceMFA™ does the work that identity verification alone cannot. It cryptographically binds a verified human’s intent to a specific transaction or action, producing a record that is genuinely authorized, provable to a regulator or auditor after the fact, non-repudiable, and immutable. Whenever US rules do arrive, and Jakimenko’s piece is a strong argument that they should, banks with an evidentiary trail of who authorized what, and when, will already be compliant with a standard that hasn’t been written yet. It’s actually creating the standard now, by default.
The takeaway for US banks reading the American Banker piece
Don’t wait for a US Annex III that isn’t even drafted. The attack Jakimenko describes is already live inside American banks today, the FinCEN red flags describe symptoms without addressing the underlying verification system, and Europe’s own timeline shows how long formal accountability takes to arrive even once regulators start drafting it. Real Human? Right Human? Right Outcome? The first two questions are table stakes. The third is what turns a compliant verification system into a defensible one, regardless of which regulator eventually shows up to check.
