Verifus and the Industrialization of Deepfake Onboarding Fraud

The full Verifus attack architecture

Verifus is rarely used alone. The standard configuration pairs Verifus with four additional components, each of which defeats a specific layer of the verification SDK's defenses.

Open Broadcaster Software (OBS). A free, open-source streaming and recording application widely used by gamers and content creators. In the fraud configuration, OBS manages the input streams and overlays the live deepfake feed on the camera output.

Virtual camera drivers. Software that exposes the OBS-managed scene as if it were a physical webcam. Most virtual camera drivers are commercial products with legitimate professional use cases, including OBS Virtual Camera, ManyCam, and XSplit VCam.

Real-time deepfake software. A generative model that produces facial movement in real time, responding to the verification SDK's prompts. The deepfake responds with the same cadence a real human would.

The Verifus Keybox. This is the component that elevates the stack. Keybox is a device fingerprint masking tool. It modifies the device telemetry that the verification SDK collects, including hardware identifiers, OS version signals, sensor characteristics, and network attributes, so that the device fingerprinting layer reports a clean, unflagged device.

That stack defeats both the camera signal and the device signal in the same session.

How it gets sold

Telegram bots, escrow services, and tiered pricing. Buyers pay in stablecoin and receive licence keys, scene templates, and pre-trained deepfake models tuned to specific document formats. Pricing tiers range from approximately 200 dollars for basic access to several thousand dollars for premium packages that include identity bundles, document templates pre-tuned for specific banks, and dedicated support channels.

This is the part that should trouble compliance teams. The technical barrier has collapsed. The fraud population is no longer constrained to nation-state actors and organized crime rings with engineering capacity. It now includes anyone who can find the right Telegram channel, transfer 200 dollars in USDT, and follow a setup tutorial. Deepfakes now account for 11 percent of global fraudulent activity.

The defensive question every CCO should be asking

If your IDV vendor has not told you, in writing, in the last 90 days, what specific controls they run against injection attacks, you have a gap. The right answer involves at least three layers.

Frame-level injection detection. The captured frames are analyzed for the artifacts virtual cameras leave behind. Driver-specific encoding patterns. Frame timing inconsistencies. Pixel-level signatures that real device sensors do not produce.

Hardware attestation at the OS level. The verification SDK demands a cryptographic attestation from the device's secure enclave that confirms the capture originated from a real camera sensor on a non-tampered OS. The Verifus Keybox cannot defeat hardware attestation.

Behavioral biometric layering during the session itself. The deepfake is too smooth. Real humans hesitate, oversteer, overshoot. The signature is detectable at sub-second resolution.

What this means for the next 90 days

The Verifus stack is going to spread. The economics are too good for it not to. Two specific recommendations.

First, run a red team against your own onboarding using publicly available injection tooling. The exercise takes two days and costs less than 2,000 dollars. The cost of not running it is several orders of magnitude higher.

Second, demand SOC 2 Type II language from your IDV vendor that explicitly addresses injection attack defenses. The absence of that language is the evidence.

The regulatory implications

The FinCEN AML/CFT Reform Notice of Proposed Rulemaking, currently open for public comment, explicitly references the obligation to maintain effective fraud prevention measures proportionate to the institution's risk profile. The UK's Failure to Prevent Fraud offence, now six months in force, requires organizations meeting the size threshold to demonstrate "reasonable procedures" against fraud committed by associated persons. In both regimes, an institution that has not addressed Verifus-class injection attacks in its fraud prevention architecture is going to have a difficult conversation with examiners.

The fraud-as-a-service economy reached a new milestone this week. Every compliance team in the world should treat that milestone as a deadline, not a curiosity.