Catching what hash databases can't see.
Visork is an ML-first classifier. Trained to recognise CSAM — including new and AI-generated material — in real time, inside your upload pipeline.
Score well above threshold — clear block.
Score 0.974 against a 0.85 threshold. Visork returns a block verdict — engineered for sub-200 ms — with a signed audit log id.
Two ways to detect harm.
Visork is the second one.
Hash matching is fast but only finds files that already exist in a database — it can't see new uploads or AI-generated material. The Visork ML classifier costs roughly ten times a hash lookup — still inside the sub-200 ms end-to-end target — and in exchange it works on content nobody has ever indexed. The two are complementary; the right answer for most platforms is to run them side by side.
Hash matching (industry)
Detects only what's been seen before. Requires access to authorised hash databases (NCMEC, IWF) which are gated.
Hash matching (industry)
Detects only what's been seen before. Requires access to authorised hash databases (NCMEC, IWF) which are gated.
ML classifier (Visork)
Detects novel and AI-generated content. Every analysis returns a probabilistic score — false positives and false negatives are first-class quality metrics, calibrated against representative data.
You set the threshold per module to fit your platform's risk profile. The verdict is a recommendation; the action — block, quarantine, review, allow — is yours.
One upload in a hundred thousand.
Take the illustrative base rate from our explainer: one upload in a hundred thousand is abuse material. At that rarity nobody finds it by looking, and a vendor's “accuracy” figure stops meaning anything — a classifier that flags nothing at all scores 99.999 %. Visork returns a probabilistic score for every image and is measured on precision and recall, on your own traffic, during the pilot's 100 free analyses.
We don't quote an accuracy number.
We report precision and recall at a stated threshold, measured on your traffic — not on a curated test set where positives are a tenth of the data and every figure collapses on contact with a real upload stream.
Why “99 % accurate” tells you nothing- Precision
- Of everything flagged, how much was actually abuse material. Your review queue feels it.
- Recall
- Of the abuse material present, how much was caught. Your legal exposure feels it.
What's live today. What's next.
The same architecture extends across modules. We ship them one at a time — accuracy first, then coverage.
Live with pilot platforms. Trained for novel and AI-generated content.
Same architecture, separate module.
Detection of AI-generated impersonation imagery.
General explicit-content detection alongside CSAM.
Conversational risk detection over chats and comments.
See it on your data.
Pilots start with 100 analyses at no cost, benchmarked against your own traffic rather than our demo set.