Frequently asked questions
Does my document really never leave my browser?
Correct, and you don't have to trust us on it. Open DevTools > Network, process a document, and watch: zero outbound requests during processing. Beyond observation, the page's Content-Security-Policy restricts every connection to 'self', the browser itself will refuse any external request. Detection, OCR, encryption, and redaction all run locally via WebAssembly.
Do we need a BAA?
No. ZeroEgress never has access to PHI at any point, there is no upload and no server-side processing, so it does not meet the HIPAA definition of a Business Associate (45 CFR § 160.103). There is no data path through our infrastructure that a BAA would govern.
Do we need a DPA? Is ZeroEgress a GDPR data processor?
No. No personal data is processed on our infrastructure, processing happens on your device, under your control. There is no processor relationship because there is no processing on our side.
Do you have SOC 2?
Not applicable, and that's the point. SOC 2 audits vendor infrastructure that handles customer data. Ours handles none. The only thing our infrastructure does is serve static files (the app, and the AI model itself) from our own origin.
What about your data breach notification policy?
A breach of our infrastructure cannot expose your documents, because your documents are never on it. The worst-case impact of compromising our servers is serving a modified application, which is why the WASM engine and the model manifest are both signed (ECDSA-P256) and verified in your browser against a pinned key before anything runs, and why every segment of the model download is SHA-256-verified before use. A mismatch stops the application rather than running modified code.
Can our auditors verify the audit trail independently?
Yes. Every session can export an ECDSA-signed audit trail (JSON or CSV), the redaction map, and a signed compliance certificate built only from what the session observed. They verify offline with standard tools, no access to our infrastructure needed, and the verification steps are documented in the export itself. A sample evidence pack is available on request.
Can you prove when the records were created?
To a defined extent, yes. At its origin the audit chain binds an external time value read from the response that delivers the AI model, so any verifier who trusts whoever issued that value learns the records cannot have been created before it, a verifiable earliest bound. Ordering within the session comes from a monotonic counter rather than the local clock, so a reordered or back-dated sequence is detectable even if the machine's clock is wrong. For a latest bound, the exported evidence bundle contains a digest of the chain head that your deployment can have timestamped externally, by a timestamping authority or an append-only log; that step is available, not something we do for you. Both directions involve only digests and time values, nothing about the documents leaves the browser.
What does it detect?
AI-based NER (names, organizations, locations) plus 17 pattern categories: email addresses, phone numbers, fax numbers, Social Security Numbers, credit card numbers, IBANs, IP addresses, URLs, dates, ZIP codes, account numbers, medical record numbers (MRN), patient and claim IDs, health plan beneficiary numbers, driver's license numbers, VINs, and national ID numbers. A second pass propagates every confirmed entity across the whole document, including ALL-CAPS variants. You can add custom patterns and whitelist terms.
How accurate is it?
The NER model (dslim/distilbert-NER) has roughly 91% F1 accuracy. That means it misses some entities and occasionally flags the wrong thing. Review the results before relying on them, the interactive review step exists precisely for that.
Which file types are supported?
PDF (including scanned, via in-browser OCR), Word, Excel, PowerPoint, images (JPG/PNG/TIFF), and plain text. Batch mode processes multiple files into a single redacted PDF with one combined audit trail.
Why is the first load ~250MB?
That's the AI model downloading, once, from our own origin (not from any third-party CDN), integrity-checked and then cached by your browser. After that, ZeroEgress works offline. The size is the price of running real AI locally instead of sending your data to someone else's.
Can I use the output with ChatGPT or Claude?
Yes, that's a primary use case. Reversible tokenization replaces entities with consistent tokens ([NAME_1]), so the AI keeps context without identities; the Reveal feature restores originals locally afterward. See the full workflow.
Is it patented?
Yes. The client-side AI processing architecture is covered by US Patent 12,665,758 (granted June 2026, 30 claims), with additional applications allowed and pending.
What does it cost?
Free during beta, no signup, no credit card. Team and enterprise plans will be announced ahead of general availability.
Can we run it on our own infrastructure?
The application is static files plus a self-hosted model, it can be deployed on internal infrastructure for fully air-gapped use. An enterprise deployment package is on the roadmap; if you want to talk about it.