Supported findings · Separate output · Human review

Document Privacy Scanner for AI Uploads

FilePreflight is being built to find supported personal data, secrets, identifiers, metadata, comments, and hidden content before a file is shared with AI.

Initial Windows release in preparationNo download yetEvidence pending

Product direction

From a finding to a checked copy—not a dashboard full of claims.

The Windows MVP focuses on the shipped path: choose a supported file, inspect it, remove selected supported items, reopen the output, check again, and give one clear next action. Advanced reports and safety certificates are not part of the MVP.

Supported risk categoriesNames, contact details, identifiers, credentials, secrets, document metadata, comments, hidden content, and selected image metadata or OCR regions.
Format-aware transformationPDF, Office, image, text, and structured-data files require different removal and verification methods.
Clear limitsUnsupported formats, edition-limit violations, failed reopening, and unresolved critical secrets stop before a safe output is shown.

What the scanner is designed to check

People and contact dataNames, email addresses, phone numbers, postal information, and user-supplied terms.
Business identifiersCustomer, order, invoice, account, contract, and employee identifiers.
Credentials and secretsAPI keys, access tokens, passwords, private keys, connection strings, and secret environment values.
Hidden file dataAuthors, company properties, comments, revisions, hidden elements, image location metadata, and supported embedded structures.

What a result can—and cannot—mean

ResultMeaningRequired next step
Safe copy createdThe output passed the supported reopen and recheck pathReview the exact output and destination before sharing
Review requiredA supported finding or file condition needs judgementInspect the summary and decide whether to change or stop
Blocked or failedThe app cannot publish a verified normal outputFollow the stated reason; the original remains unchanged

A successful technical result does not replace organisational policy, consent, legal review, or destination governance.

Evidence before broad claims

Coverage must be demonstrated with synthetic fixtures and format-specific checks. Local-processing claims need a documented network-behavior method. Conversion claims need outputs that reopen, preserve useful content, remove the selected canaries, and leave the original unchanged.

The evidence hub separates planned methods from verified release results so pre-release copy does not imply a test has already passed.

Evidence and next-step resources