You Have Been LaTeXpOsEd: A Systematic Analysis of Information Leakage in Preprint Archives Using Large Language Models
This paper presents "LaTeXpOsEd," a systematic large-scale security audit of over 100,000 arXiv submissions that reveals widespread information leakage in source files and comments, including sensitive credentials, PII, and confidential communications, by leveraging a novel framework integrating pattern matching and large language models.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are submitting a manuscript to a prestigious library (like arXiv) to share your latest scientific discovery. You hand over the final, polished book (the PDF) that everyone can read. But, because the library wants to ensure your work is honest and reproducible, they also ask you to hand over your entire workshop: the rough drafts, the sticky notes on the walls, the half-finished sketches, and the notes you scribbled in the margins while thinking out loud.
The paper "You Have Been LaTeXpOsEd" is a massive security audit of this workshop. The researchers found that while scientists are very careful about what they put in the final book, they often forget to clean out their workshop before handing it over.
Here is a simple breakdown of what they found, using everyday analogies:
1. The "Digital Attic" Problem
When scientists upload their papers, they don't just upload the final PDF. They upload the LaTeX source files. Think of these source files as a digital attic or a garage.
- The Mistake: Authors often leave "sticky notes" (comments in the code) or "old boxes" (unused files) in this attic.
- The Risk: These notes might say things like, "Hey, the password for our internal server is 'Password123'" or "We had a huge fight with the reviewer, here are the angry emails."
- The Reality: Because the attic is public, anyone can walk in and read these notes. The researchers treated this as a "passive" audit—they just looked at what was already there, like a security guard walking through an open house, without trying to break in or test if the locks actually worked.
2. The Detective Tools: Regex vs. The "Super-Reader"
To find these secrets, the researchers used two types of detectives:
- The Pattern Matcher (Regex): This is like a robot that only looks for specific shapes. It can easily find things that look like an email address (
name@domain.com) or an IP address (192.168.1.1). It's good at finding obvious clues but bad at understanding context. - The "Super-Reader" (Large Language Models - LLMs): This is like a highly intelligent human detective who can read a sentence and understand the meaning.
- Example: A pattern matcher might miss a sentence that says, "I'm using the key 'SecretKey99' to log into the lab." because it doesn't look like a standard password format.
- The LLM, however, reads the whole sentence, understands that "SecretKey99" is a secret, and flags it.
The researchers tested 25 different "Super-Readers" (AI models) to see which one was the best detective. They found that open-source models (like Qwen-2.5) were almost as good as the expensive, proprietary ones, but cost a fraction of the price.
3. What Did They Find in the Attic?
After scanning 100,000 papers (about 1.2 terabytes of data, which is like a massive library), they found thousands of "leaks." Here is what was left behind:
- The "Open Door" Keys: Hundreds of instances of actual passwords, API keys, and login credentials. It's as if someone left their house key taped to the front door with a note saying, "This opens the back door."
- The "Private Chat" Leaks: Links to private Google Drives or Dropbox folders that were set to "Anyone with the link can edit." This meant that confidential peer reviews, internal arguments between co-authors, and unpublished data were visible to the whole world.
- The "Sticky Note" Arguments: Comments where authors were arguing about the quality of the research or complaining about reviewers. These were meant for internal eyes only but were published for everyone to see.
- The "Hidden Photos": Images uploaded to the paper often contained EXIF data (digital metadata). This is like a photo of a building that secretly includes the GPS coordinates of the photographer's home and the exact time the photo was taken.
4. The Cost of the Investigation
One of the most surprising findings was how cheap it was to do this.
- The researchers spent about $90 total to scan 100,000 papers.
- This included the cost of downloading the data and paying the AI to read the comments.
- The Analogy: It's like hiring a security team to check 100,000 houses for a few dollars. This means that even a "low-resource" bad actor (someone with a small budget) could easily do this to steal secrets or cause reputational damage.
5. The "Blame Game" (Who is Responsible?)
The paper points out a tricky legal situation.
- The Library's Rule: When you submit to arXiv, you sign a contract saying, "I grant the library the right to distribute my work forever." The library also says, "It is your job to make sure you don't include private info."
- The Problem: Authors often think, "I put the secret in the code comments; no one reads the code, only the PDF." They don't realize that the "code comments" are part of the permanent public record.
- The Verdict: The researchers argue that while the authors are technically responsible for cleaning up, the system is broken because it's too easy to accidentally leave secrets behind, and the library doesn't automatically scrub them.
6. The Solution: Cleaning the Workshop
The paper suggests two main fixes:
- For Authors: Before hitting "Submit," run a "clean-up" tool (like a digital vacuum) that automatically deletes all comments, unused files, and hidden metadata. Don't leave sticky notes in the attic!
- For the Library (arXiv): They should add a safety net. Before a paper goes live, the system should automatically scan for passwords and private links and warn the author: "Hey, you left a password in your notes. Do you really want to publish this?"
Summary
The paper is a wake-up call. It shows that in the rush to share science quickly, researchers are accidentally leaving their "digital trash" (secrets, passwords, and private arguments) out in the open. Because modern AI is so good at reading and understanding these messy notes, it's now incredibly easy for anyone to find them. The researchers urge the scientific community to start cleaning up their source files before publishing, just as they would clean up a physical lab before opening the doors to the public.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.