Authenticity examination

Know what’s real
before you act on it.

Safe Trust AI examines video, audio, images and documents and tells you whether they were captured or synthesized — with the evidence trail your risk, identity and trust & safety teams need to stand behind the call.

Operated by Safe Trust Lab AI Pte. Ltd. (Singapore · UEN 202615952G) and Safe Front, Inc. (Delaware C-Corp, United States) · Member of NVIDIA Inception

Built for the teams that carry the consequences of getting authenticity wrong.

Financial servicesIdentity & KYCInsurance & claimsNewsroomsPublic sectorMarketplaces
01 The exposure

Synthetic media stopped being a novelty. It became an attack surface.

The same generative tools that make a demo feel impressive are now used to move money, open accounts, and manufacture evidence. The question is no longer whether your team will meet a fake — it’s whether you’ll catch it in time.

$25M

One call can clear a wire

A single video meeting populated with cloned executives was enough to authorize a multi-million-dollar transfer. Voice and likeness are no longer proof of presence — and your controls were written for a world where they were.

// Reported finance-sector incident, 2024
10×

Fraud is scaling faster than review

Detected face-swap and voice-clone attempts against identity systems have multiplied year over year — faster than manual review teams can absorb.

// Directional industry estimate
Mins

A convincing fake is now cheap

What used to need a studio needs a laptop and a reference clip. Defenders need examination that keeps pace with how fast generation moves.

// Open-source tooling availability
02 Why teams trust the verdict

A score is an opinion. An evidence trail is a decision you can defend.

Anyone can return “real” or “fake.” What lets your team act on it — and answer for it later — is being able to show the work behind every call.

VALUE 01

Calibrated confidence, not a gut feeling

Scores are tuned so that a 90 means the same thing across millions of files. Your thresholds stay meaningful over time instead of drifting as content changes.

  • Set distinct thresholds for auto-action vs. human review
  • Consistent meaning so your policy doesn’t quietly erode
Signal weightingvideo · face
Facial landmark consistency96
Blink & micro-expression timing93
Lighting & shadow coherence91
Frame-to-frame warping74
Lip-sync alignment94
VALUE 02

Every verdict shows its working

See which detectors fired and how heavily each one counted, exported as a signed, tamper-evident report — so a flag is explainable to a regulator or counsel, not just a red light on a dashboard.

  • Per-signal breakdown with provenance and reviewer trail
  • Exportable for case files, disputes, or disclosure
Report · STA-7F3C9flagged
87.4%
Likely synthetic
Above review threshold (75%)
source.sha256  a3f9…e21b
video.blink_rate  irregular ⚠
video.lighting  inconsistent ⚠
audio.spectral_seam  detected ⚠
audio.prosody  within range ✓
metadata.c2pa  absent
reviewer  k.chen · 2026-06-18T09:14Z
VALUE 03

One examination across every medium a fake hides in

Attackers don’t stay in one format, so examination can’t either. The same platform inspects faces, voices, images and documents — and feeds verdicts into the tools your team already works in.

  • API into your fraud stack, or a console for analysts
  • Identical auditable record from either path
See it in production
Media coverage4 types · 20 signals

Video & face

Face swaps, reenactment, synthetic avatars

6 signals

Voice & audio

Cloning, splicing, text-to-speech

5 signals

Images

Diffusion-generated, inpainted, edited

5 signals

Documents & IDs

Forged IDs, tampered PDFs, fabricated records

4 signals
03 How to start

From first upload to a verdict you can defend.

Every file moves through the same path. Nothing is a black box you have to take on faith — each stage leaves something you can inspect.

STEP 01

Bring your hardest cases

Start with the files that fooled your team — or nearly did. No production data or integration is required to run a first evaluation under NDA.

~1 day
STEP 02

Connect via API or console

Pipe media in through the API to your existing fraud and case-management flow, or let analysts upload directly in the review console. Both paths produce the same record.

~1 week
STEP 03

Examine & weigh signals

Independent detectors look for the artifacts generation leaves behind — irregular blink patterns, spectral seams, metadata mismatches. Signals are weighed, not blindly averaged.

<3s / clip
STEP 04

Act on the verdict, keep the trail

You get a calibrated score, a plain verdict, and a per-signal report your auditors, regulators or counsel can read — set your own thresholds for what acts automatically.

ongoing
04 In production

What changed for teams already using it.

Engagements are described with sector and outcome; client names are withheld under NDA during this stage.

Onboarding · video KYC
Digital bank · Southeast Asia

Catching cloned-voice verification at onboarding

ProblemA wave of account-opening attempts used cloned voices to pass live verification that human reviewers were waving through.
ApproachVoice and video examination wired into the onboarding flow via API, with high-confidence fakes auto-routed to manual hold.
ResultA full batch caughtFraudulent applications flagged in the first onboarding cohort — enough to justify the rollout on its own.
Trust & safety
Online marketplace

Triage before a human opens the file

ProblemEvery disputed video was reviewed manually, burying analysts in obvious fakes.
ApproachAutomated examination flags clear synthetics up front; ambiguous cases reach analysts with evidence pre-assembled.
ResultAnalyst time reclaimedRedirected from obvious fakes to genuinely ambiguous cases.
Compliance sign-off
Insurance & claims platform

Evidence the compliance team could read

ProblemAccuracy alone couldn't win internal sign-off for automated decisions.
ApproachPer-signal, exportable reports gave compliance a trail they could stand behind in a dispute.
ResultInternal sign-off securedWon on the strength of the evidence trail, not the score alone.
05 Before you ask

Still have a question?

We’d rather show you the real state of things than overstate a badge. Bring your hardest files and we’ll walk through the evidence together.

Talk to the team

Examination runs in regional infrastructure you select, on a data-minimization basis — files are processed to produce a verdict and evidence record, with retention controlled by your policy. Dedicated and on-premise deployments are available for teams that cannot send media off their own environment.

We don't claim to be infallible — that's exactly why every verdict ships with a confidence score and a per-signal breakdown. You set the threshold for automated action versus human review, and ambiguous cases route to analysts with evidence already assembled. Fewer fakes slip through; fewer real files get wrongly blocked.

Detection is an adversarial, moving target, so we treat it as one. Signal models are updated as new generation techniques emerge, and because we rely on multiple independent detectors rather than a single classifier, no one new technique blinds the whole system at once.

Most teams integrate via API and pipe verdicts straight into their case management or fraud-decisioning flow. Analysts who prefer a workspace use the review console. Either path produces the same auditable record, so you don't have to choose between automation and a paper trail.

We're transparent about what is in place now versus on the roadmap. Security and compliance documentation can be reviewed during vendor assessment — we'd rather walk you through the real state of controls than overstate a badge.

Request access

Run your hardest cases through it.

Bring the files that fooled your team — or nearly did. We’ll set up an evaluation against your real media and show you the evidence behind every verdict.

// Evaluations run under NDA. No production data required to start.