Visible sensitive content
Names, email addresses, phone numbers, account identifiers, contract terms, customer details, and other information shown on the page.
Local-first Windows app · Initial release in preparation
AI Document Sanitizer: Make Files Safe Before You Upload
FilePreflight is a Windows app in development for inspecting supported sensitive content, hidden metadata, and risky file elements on-device—then exporting a separate copy for you to review before an AI upload.
What “safe” means here: reducing identified exposure in supported file types—not a guarantee that a file is anonymous, compliant, or risk-free. Automated detection is best-effort, so review the exported copy before uploading.
Initial Windows release in preparation. No download or checkout is available yet.
customer-analysis.pdfSeparate output copy planned
Visible contentNames, email addresses, account details
ReviewHidden file dataComments, notes, embedded elements
InspectMetadata & secretsProperties, tokens, configuration values
ScanExport and recheckUser review remains required
01 Local-first design
02 Separate-copy workflow
03 Best-effort detection
04 Human final review
Before ChatGPT, Claude, Gemini, or another AI service
A file can reveal more than the page you see.
Copying a document and deleting a name may leave other clues behind. A useful AI document sanitizer must separate what is visible from what can be stored inside the file.
Hidden document data
Comments, tracked changes, hidden text, hidden sheets, speaker notes, embedded files, and other format-specific content outside the normal view.
Metadata and secrets
Author and company properties, document history, image metadata, API keys, tokens, passwords, connection strings, and configuration values.
Microsoft documents hidden Office data that may require inspection, while Adobe distinguishes visible PDF redaction from hidden-data sanitization. See the primary sources.
A reviewable preflight, not a magic “safe” badge
Inspect. Decide. Export. Recheck.
The planned workflow keeps the decision with the person who understands the file and the destination.
- 01
Start with the task
Decide what the AI actually needs. Remove entire sections, sheets, or fields that are unnecessary before relying on pattern detection.
- 02
Inspect supported risks
Review visible content, file properties, hidden elements, and supported credential patterns using checks appropriate to the format.
- 03
Choose each change
Remove, replace, or generalize selected findings. Context matters, so the software should assist your judgment rather than silently decide for you.
- 04
Export and recheck
Write a separate copy, reopen it, confirm it is readable, scan it again, and review that exact output before uploading.
Evidence over slogans
Proof should be inspectable.
The release will be accompanied by artifacts that let you evaluate the local-first workflow and its limits. Until they are published, this page describes those capabilities as planned.
Evidence package: publishing with the initial release
Network behavior test
A dated, reproducible test with the environment, method, and observed network behavior documented.
Synthetic before & after files
Downloadable examples containing deliberate visible and hidden test data, paired with the exported copy.
Format-and-limits matrix
Per-format coverage for inspection, transformation, rechecking, and cases that still need manual work.
Example review record
A sample findings and recheck summary—presented as a review aid, never as security certification.
Planned inspection layers
One workflow for four different exposure modes.
Each file format stores information differently. FilePreflight is being designed to apply supported, format-aware checks and show what still needs review.
Sensitive information
Supported patterns for personal and business identifiers in visible or extracted content.
Context-sensitive facts may not match a rule.Hidden Office & PDF data
Supported comments, notes, hidden content, attachments, and other embedded elements.
Coverage varies by format and file structure.Metadata
Supported author, company, title, timestamp, image, and document-property fields.
Some properties may be stored outside tested fields.Credentials
Supported API-key, token, password, connection-string, and configuration patterns.
If a secret was exposed, revoke or rotate it.Planned initial Windows coverage
Formats and limits, side by side.
The 17 extensions below are targets for the initial release, subject to reproducible verification. The public support matrix will replace these planned labels at launch.
| File group | Planned extensions | Inspection focus | Important limits |
|---|---|---|---|
| Documents & Office | .pdf .docx .xlsx .pptx | Visible text, supported metadata, comments, notes, hidden or embedded elements | Encrypted files, macros, linked content, complex objects, and formula effects may require manual work |
| Images | .jpg .jpeg .png | OCR text and supported image metadata | OCR depends on image quality, layout, language, and handwriting; visual review remains essential |
| Text | .txt .md .markdown .log | Supported identifiers, credentials, and text patterns | Project names, commercial sensitivity, and uncommon secret formats can require contextual review |
| Data & config | .csv .json .xml .yaml .yml .env | Structured values, identifiers, keys, tokens, passwords, and connection details | Relationships between fields can remain identifying even after direct identifiers are removed |
No complete-detection guarantee. “No findings” means only that no supported rule found a match in the tested scope.
No automatic compliance. Your authorization, destination, provider settings, contracts, policies, and applicable law still matter.
No substitute for review. Re-identification and inference risks can remain after direct identifiers are removed.
Local-first by design
Prepare the copy before it reaches an AI service.
The planned application workflow performs file inspection and transformation on the Windows device. Once you upload the reviewed copy to ChatGPT, Claude, Gemini, or another service, that copy leaves the device and the provider's terms and settings apply.
1Original filePreserved under your controls
2Local preflightInspect, transform, recheck
3Reviewed copyYou decide whether to upload
Planned perpetual pricing
Start free. Upgrade for workload and control.
Free and Pro are planned to use the same per-file processing approach. Published features, limits, pricing, and availability may change before release.
Build the full workflow
Prepare the file—and check the destination.
A cleaner copy is one layer. Use these source-backed guides to review the AI service, the document type, and the disclosure decision.
Decision guideIs it safe to upload confidential documents to ChatGPT?Read guide →
RedactionRedact sensitive information before ChatGPTRead guide →
Company dataAnonymize company data before AIRead guide →
SpreadsheetsAnonymize Excel data before AIRead guide →
Hidden dataRemove hidden data and metadata before AIRead guide →
PDFRedact and sanitize a PDF before ChatGPTRead guide →
Frequently asked questions
Know the boundary before you rely on the tool.
FilePreflight is a preparation aid. The person sharing the file remains responsible for the final review and upload decision.
What does “safe” mean here?
Safe means reducing identified exposure in supported file types before upload. It does not mean anonymous, compliant, or risk-free. Detection is best-effort, so review the exported copy.
Do my files leave my computer?
FilePreflight is being designed for on-device file processing. The release evidence will document network behavior under stated test conditions. Until that evidence is published, local processing is a product design goal rather than an independently verifiable claim.
Does FilePreflight modify the original?
The planned workflow creates a separate output copy and leaves the source file unchanged. Preserve the original, then review and re-inspect the exported copy before sharing it.
Is redaction the same as sanitization?
No. Redaction generally removes visible text or graphics. Sanitization also addresses supported hidden data such as metadata, embedded content, comments, or scripts. Some files need both.
Can an automated scanner find every sensitive item?
No. Detection depends on file type, structure, content, selected rules, and context. A result with no findings is not proof that a file contains no confidential information.
What about scanned PDFs and complex spreadsheets?
Scanned pages may depend on OCR quality. Hidden sheets, formulas, external links, macros, embedded objects, and encrypted content need format-specific handling and may require manual review. Final support will be documented per format.
Does sanitizing a file make an upload compliant?
No. File preparation can support data minimisation, but compliance depends on your purpose, authorization, policies, contracts, provider settings, security controls, and applicable law.
Primary sources
Why these layers matter.
Microsoft documents comments, revisions, document properties, hidden text, hidden spreadsheet rows and columns, and other information that may remain in Office files. It also describes items its Inspector cannot detect or remove and recommends inspecting a copy.
Adobe distinguishes visible-content redaction from sanitization of hidden information such as metadata, embedded content, and scripts. NIST explains that de-identification techniques vary in effectiveness and that re-identification or inference risk can remain.
FilePreflight for Windows
Make the upload decision with a file you have actually reviewed.
Join the launch list for availability, verified format coverage, and the first public evidence package.