The Epstein files: what a huge release teaches about redaction at scale

Between December 2025 and early 2026, the US Justice Department published millions of pages of files connected to Jeffrey Epstein. The releases became one of the most closely examined redaction projects ever. Reporters, lawmakers and survivors' lawyers found problems in both directions: information that should have been hidden wasn't, and information that didn't need hiding was.

This post isn't about what the files say. It's about how the redaction went wrong, because the same failures can happen in any firm handling sensitive documents under a deadline.

What went wrong

Covered, not removed. Shortly after the first release in December 2025, people found that in some files the black boxes could be bypassed by copying the text and pasting it elsewhere. Mediaite reported that "redacted" passages in some documents were readable this way. Not every file was affected, but those that were had the same flaw as the court filings we've covered before: the box was drawn on top, and the text stayed underneath.

Identities left visible. In early 2026, lawyers for survivors told the court about what they described as thousands of instances where names and other identifying details hadn't been redacted. NPR reported that the department attributed the problems to "technical or human error," took down several thousand documents and media files, and said it had revised how it handles flagged documents, aiming to repost corrected versions within 24 to 36 hours. Reporting also noted that one email in the release showed a full credit card number, expiry date and security code.

The same document, redacted differently. NPR found that one presentation appeared six times across the files, with different information blocked out in each copy. Separately, first and last names sometimes appeared in different documents, so combining them made people identifiable.

Too much hidden, too. CBS News found that more than 500 pages were entirely blacked out in the initial release, and after lawmakers reviewed unredacted versions, the department un-redacted further names it said were not victims.

Lessons for any firm

The scale is unusual. The mistakes aren't.

1. Removal has to be the default. If the tool can draw a box without removing the text, someone eventually will. Use a process where every export is clean by construction, not one that relies on remembering an extra step.
2. Identifiers are more than names. Email addresses, nicknames, phone numbers, faces in photos and card numbers all identify people. Build your checklist around everything that could point to a person.
3. Duplicates must match. When the same document appears more than once, redact it once and reuse that decision. Redacting each copy by hand invites the six-different-versions problem.
4. Think across documents, not just within them. A first name here and a surname there can be enough. Reviewers need to know what's already public in the same release.
5. Volume needs process, not heroics. Large productions on tight deadlines are exactly when ad hoc redaction breaks down. Plan reviewer time, sampling and a final check before anything is released.
6. Have a way to fix mistakes fast. The department's 24-to-36-hour takedown-and-repost process came after the fact. Decide in advance who can pull a document and how quickly.

Where Ombly fits

Ombly is designed around the first three lessons. Every export is rebuilt from pixels, so nothing covered stays underneath. Detection looks for names, emails, phone numbers, card and account numbers and your own patterns, and a person reviews each suggestion. For teams, it runs on your own servers and records what was removed and when, without storing the sensitive text. Book a call to see it with your own documents.