How PrivScrub works — methodology
File processing happens in your browser. This page explains how supported PDFs are changed, what is checked in the generated output, and what still needs your review.
Page updated 2026-07-08
Supported PDF scope
PDF redaction and metadata removal accept basic, unencrypted PDFs up to 50 pages and 25 MiB. Nonempty annotations, internal or external links, forms, attachments, actions and layers are unsupported. A detected unsupported structure stops the workflow before export.
Redaction rebuilds every page
Every output page is rebuilt as an image in a new PDF, with selected regions covered in its pixels. Unboxed areas retain their visible appearance, but every page loses searchable and accessible text. No original pages or resources are copied. Review every output page visually; this workflow does not run OCR.
The actual output is reopened
After creating the new file, PrivScrub rereads those bytes to check extractable text, supported document Info values, XMP presence and PDF structures. The result identifies the output using its actual SHA-256, size and page count. A hash identifies the file; it is not proof that the contents are safe.
Exact text checks are scoped
Optional target strings are searched in extractable text and supported document Info values, using case-sensitive matching with normalized whitespace and no cross-page matches. XMP contents and image text are not searched. Targets do not automatically select redaction regions. They remain in current session memory and are excluded from downloaded check records. No target supplied, incomplete checks and findings are stated separately.
Metadata removal is a separate operation
The metadata tool clears document Info fields and XMP while retaining supported page content. It then rereads its own generated output. Removing metadata does not redact visible names, addresses or images. Unsupported structures stop processing before export.
Other privacy tools
Image tools and PII inspection have their own limits. Pattern and checksum matches identify candidate personal data, not confirmed identity. Image masking and face detection still require review for missed regions. The specific tool explains its supported operation.
What the result does not establish
Passing checks apply only to their named scope. They do not certify that all sensitive details are gone, examine text inside images with OCR, or prevent information being inferred from the remaining document. Open the downloaded PDF and visually review every page before sharing.
Local processing and network use
File processing runs locally; file contents are not sent to a processing service. Loading the site and its required resources can use the network. Once required resources have loaded, supported processing can continue offline.