2026-09-05

Redacting a PDF and stripping its metadata are two different jobs — NearScrub only does one

Two very different operations get described as "cleaning a document before you send it." One is redaction: taking out information that is on the page — a name, an address, a paragraph you are not permitted to disclose. The other is metadata removal: clearing the fields around the page — author, producing software, creation and modification timestamps, the XMP block — none of which appear anywhere when you read the file. Different data, different failure modes, different tools. NearScrub does the second one only. That is worth stating plainly, because the two get conflated most often in exactly the settings where being wrong about it is expensive: court filings, FOIA responses, disclosures, anything published as a PDF that someone else has a reason to pull apart.

A black box drawn on a page is not a redaction

The best-known demonstration is still the January 8, 2019 filing by Paul Manafort's defence lawyers. As Columbia Journalism Review recounted, Guardian reporter Jon Swaine found that copying the blacked-out passages and pasting them elsewhere made the bars disappear and left the text behind — revealing the Kilimnik meetings and the shared polling data the filing was meant to keep sealed. No forensic tooling was involved. The black rectangles were drawn objects sitting on top of text that was still in the file, still selectable, still copyable.

Courts have been saying this in writing for a long time. The U.S. District Court for the District of Massachusetts published an Updated Notice Regarding Metadata and Redacted Information in PDF Documents on January 11, 2010, with a list headed "Methods Not to Use." It names changing the font to white ("this only appears to make the words disappear"), and it names "Use of Adobe Acrobat graphic and commenting tools that can black-out, cover over or remove sections of text," with the parenthetical "these edits can be removed by anyone to reveal the text underneath." It even covers paper: ink-marking or semi-translucent tape over a page you are about to scan, because "your covered data may still show through." Its positive advice is one sentence: "The best way to redact a document is to not include the unwanted text initially."

One caveat about that notice, in fairness to it: parts are dated. Its claim that Adobe Acrobat "by itself cannot redact a document using any of the built in tools" was true of the versions in front of the clerk in 2010 and is not true of Acrobat today, which ships a redaction tool that actually removes content. The structural point — that a drawing tool and a redaction tool are different things, and that only one of them deletes anything — has not aged at all.

Even a redaction that really deletes the text can leak

Here is the part that surprises people who already know the copy-paste trick. In Story Beyond the Eye: Glyph Positions Break PDF Text Redaction (PoPETs 2023), Maxwell Bland, Anushya Iyer and Kirill Levchenko show that redactions which do excise the text can still leak it, because the surrounding characters were laid out around the words that used to be there. Sub-pixel horizontal shifts in the positions of the remaining glyphs carry enough information to recover what was removed. They assessed 11 popular redaction tools, Adobe Acrobat among them, and found all of them leaking; they then de-redacted hundreds of real-world redactions, including ones in Office of Inspector General investigation reports and FOIA responses, and notified the OIG, the Free Law Project, PACER, Adobe, Microsoft and the Department of Justice, releasing open-source fixes alongside the paper.

The honest scope of that result: the demonstrated attack recovers short, guessable strings — first and last names, tested against a dictionary of candidates to see which one reproduces the same layout — not arbitrary redacted paragraphs. It is a reason to treat a redacted name as weaker protection than it looks, and a reason to prefer the court's advice of never letting the text into the published file at all. It is not a claim that every redaction in every PDF is readable.

The other half: the fields nobody looks at

Redaction and metadata are treated as one topic in that Massachusetts notice for a reason — it is titled "Metadata and Redacted Information," and its further-reading list points readers at the NSA's own PDF redaction guidance. A file can be redacted perfectly and still announce who wrote it, on what software, at what time, in fields that no reader ever opens.

How common that is has actually been measured. In Exploitation and Sanitization of Hidden Data in PDF Files (ACM IH&MMSec 2021), Supriya Adhatarao and Cédric Lauradoux collected 39,664 PDF files published on the websites of 75 security agencies across 47 countries and recovered sensitive data from 76% of them — author names, the producing tool, operating-system details. Only seven of the 75 agencies were sanitizing their PDFs at all, and by the authors' count only three of those were removing everything. These are organisations whose job description includes not leaking information.

What NearScrub actually changes in a PDF

Exactly the bottom row. It deletes the Info-dictionary entries — Title, Author, Subject, Keywords, Creator, Producer, CreationDate, ModDate — and deletes the XMP metadata stream from the document catalog if one is present, which matters because the two are separate stores and a file can carry the same author name in both. All of it happens in the browser tab; the file is never uploaded anywhere.

What it does not do, stated plainly

It is not a redaction tool and does not pretend to be. It does not read, alter or remove page content, so text sitting under a drawn black rectangle survives the scrub intact and still copy-pastes out afterwards. It does not flatten or delete annotations and markup. It does not touch form fields, embedded file attachments, optional-content layers, or the invisible OCR text layer behind a scanned page. It cannot do anything about the glyph-positioning leak above, which lives in the page's own layout. And because the PDF is re-serialised on the way out, the output is a rewritten file: the bytes are laid out differently and the checksum will not match the input, so any workflow that compares hashes needs to expect that. Format coverage is JPEG, PNG, PDF and Office documents — HEIC and video files are not supported at all.

An order of operations that works

First, follow the court's advice where you can: take the text out of the source document and publish a fresh PDF, so the sensitive string never exists in the published file. Second, when you must redact an existing PDF, use a tool whose redaction genuinely excises content, then verify it yourself — select the whole document, paste it into a plain-text editor, and full-text search for the string you removed. That thirty-second check is precisely what would have caught the 2019 filing. Third, for scanned pages, take the Massachusetts notice's warning seriously: covering material on paper before scanning can show through, so look at the resulting PDF before it goes anywhere.

Only then clear the properties, and do it last. This ordering is not arbitrary: re-exporting, re-saving or re-printing a PDF after a scrub will stamp a fresh Producer, Creator and CreationDate back into the file. Metadata removal has to be the final step before the file leaves your machine, or it gets quietly undone by the next tool in the chain. And the same notice's own closing disclaimer is worth borrowing — the clerk's office wrote that its staff "neither endorse nor guarantee any specific method" of sanitising a document. Neither does this post. Check the file you are about to send, in the state you are about to send it.