Deepfake KYC Penetration Testing
You know your vendor's spoof detection rate. You do not know your own.
An authorized technical assessment that submits synthetic faces, generated identity documents, and synthetic voiceprints through your live onboarding pipeline: liveness, document verification, face match, video KYC, and voice authentication. The deliverable is a measured acceptance rate per vector and per document type, with the specific submissions that passed and remediation mapped to your thresholds. Fully managed and external, nothing installed.
Built for fraud, onboarding, and financial crime teams at banks, fintechs, exchanges, and payments companies.
Liveness detection bypass testing, and five more surfaces
Run the full set for a complete picture of the onboarding funnel, or scope down to the single KYC deepfake bypass you are least sure about.
Synthetic faces submitted against passive, active, and challenge-response liveness. The question is not whether the check rejects a printed photo. It is whether it rejects a rendered face that blinks, turns, and follows the prompt it is given.
Generated identity documents across the types you accept, submitted through your capture and authenticity pipeline. Acceptance is reported per document type, because a stack that is solid on passports is often not solid on residence permits.
A synthetic document paired with a synthetic selfie generated to match it. This is where synthetic identity fraud actually lives, because when both artifacts are produced together they are consistent in a way a stolen document and a stranger's face never are.
A conversational avatar submitted into your video onboarding session. It responds to prompts, turns on request, and holds a conversation with the agent or the automated flow, which is the exact condition the manual review step was never designed to catch.
Synthetic voiceprints submitted against your voice biometric enrolment and verification, including the call centre path. Tests both the automated match score and the step-up challenge that is supposed to sit behind a marginal result.
Media delivered through a virtual camera or injected into the capture pipeline rather than presented to a physical sensor. Presentation controls assume a camera is looking at something. Injection removes that assumption, and most stacks have nothing at this layer.
eKYC deepfake testing: their lab versus your stack
A published detection rate is a statement about a lab. Your acceptance rate is a statement about your configuration.
Their dataset, or this month's tooling
Vendor benchmarks run against public research datasets that generation tooling moved past some time ago. We generate artifacts with what is available now, which is also what someone opening a fraudulent account this week has available.
Default settings, or your settings
The published figure describes the product at reference configuration. You run it at thresholds tuned for your conversion rate, with rules layered on top and a manual review queue behind it. That combination has never been measured, and it is the only one that matters.
A rate, or the submissions that passed
A percentage tells your fraud team something happened. The artifacts that were accepted, with the confidence scores they came back with, tell them where the boundary sits and which way to move it.
How the assessment runs
The onboarding-specific variant of our deepfake penetration testing method, scoped to KYC and identity verification. Same discipline on authorization, generation, and evidence, pointed at the account opening funnel.
This page covers the KYC and onboarding scope specifically. For the broader service, including fraud detection tooling validation and controls outside the account opening funnel, see deepfake penetration testing. Where the target is your people and process rather than your verification stack, that is deepfake simulation.
What an identity verification red team hands back
Four layers, each one usable by a different person on your team without translation.
Acceptance rate, per vector and per document type
Not one headline number. Liveness separated from document authenticity, face match separated from both, and each document type reported on its own, because coverage is almost never even across the set you accept.
The submissions that passed, with their scores
Every accepted artifact, the confidence score the stack returned for it, and the decision path it followed. Your team can pull the same records on their side and reproduce the result rather than take our word for it.
Remediation as configuration, not procurement
Threshold changes with the conversion impact estimated, rule adjustments, and the vectors where you have no control at all and adding a layer is the only answer. Where the fix is a setting, we say which setting.
Third-party assessment documentation
Methodology, authorization record, scope, dates, and results in a form that satisfies a request for independent testing of identity controls. Written to be handed over without a rewrite.
Evidence your supervisor will accept
Customer due diligence obligations require you to have effective identity verification, and effectiveness is increasingly read as something you can demonstrate rather than something you can assert from a vendor datasheet. An independent assessment gives you a dated, documented measurement of your own controls against current synthetic media, run under authorization, with the method and the scope on record. That is the artifact underwriters, external auditors, and supervisory reviews ask for, and it is considerably easier to produce before someone asks than after. No client names appear in our public material, and the report is yours alone.
How biometric verification penetration testing stays clean
A test that pollutes your onboarding data or your fraud models is not worth the finding.
Staging first, production only if it must be
Most assessments run against a staging environment configured to match production. Where a setting only exists in production, we test a ring-fenced segment at low volume with your fraud team watching in real time.
Every submission tagged and reconciled
Each artifact carries an identifier agreed with your team before submission, so the full set can be located, matched against our log, and removed. Nothing is left behind for your models to learn from.
No real identities, ever
Faces are synthetic and correspond to no living person. Document details are fictitious. No customer record, no employee likeness, and no third party system is involved at any stage of the assessment.
Regulated identity, at volume
Organizations making identity decisions remotely, where a false accept is a loss event and a regulatory one.
Real testing, real findings
What a cloned voiceprint did to a verbal verification control, and what it revealed about the step-up path sitting behind it.
Read the case study Agentic AIAutonomous agents driving synthetic media generation and submission end to end, with no human operator in the loop.
Read the case studyIf your exposure sits in the call centre rather than the onboarding funnel, start with AI vishing simulation, or see how the channels run together in the OSES simulation platform.
Common questions
Is this legal?
Yes. It is an authorized assessment of your own systems, run under a signed agreement with named approvers, an agreed environment, agreed submission volumes, and documented abort conditions. Every submission is tagged and reconciled so your team can identify and remove it afterwards. No third party system is submitted to, and no real customer record is created.
Do you use real people's likenesses?
No. Faces are fully synthetic and correspond to no living person. Identity documents are generated with fictitious details against your accepted document types. Voiceprints used in this assessment are synthetic rather than cloned from a real customer. Nothing derived from an actual identity is submitted.
Which verification vendors have you tested against?
We do not publish vendor names or per-vendor results, in either direction. What matters for your assessment is that the method is vendor neutral: we test the stack you have running, in the configuration you have it set to, including any orchestration layer or manual review step sitting behind it.
Does this touch production?
Usually not. Most assessments run against a staging environment configured to match production, which gives clean results with no operational exposure. Where a configuration only exists in production, we test a ring-fenced segment at low volume with your fraud team monitoring, and every submission is tagged so it can be reconciled and removed.
What does the report contain?
Acceptance rate per vector and per document type, the confidence scores returned for accepted submissions, the specific artifacts that passed so your team can reproduce the result, the vectors your stack has no coverage for at all, and remediation mapped to threshold and configuration changes rather than a recommendation to buy something.
Find out what your onboarding stack accepts
Thirty minutes. We will walk your verification flow and identify which layer is worth submitting against first.
Or read the broader service page on deepfake penetration testing.
support@breacher.ai
