How to use AI on sensitive files without uploading

TL;DR

To use AI without sending private data, create a sanitized copy of the document that removes personal details and sensitive items, then run the model on that copy. Define protected terms so the AI keeps needed context while hiding sensitive data, and rebuild a safe working document instead of editing the original. This enables useful AI work while reducing privacy risk.

Transcript

This week I used AI on a file I would never ever upload cuz it was way too sensitive and the model never saw the original file. It still found the three assumptions most likely to break my launch after I removed all the sensitive stuff. The customer name, the home address, private medical note that was in there, an API key, and unreleased price. Th... Read More

Key Insights

• Airlock demonstrates a practical method to define protected terms before running AI on documents, enabling context without exposing sensitive data.
• The core idea is to separate what a task needs from what a file contains, keeping personal information out of the AI process.
• Building a separate, sanitized copy preserves the original and minimizes leakage while allowing useful AI analysis.
• Redaction alone is insufficient; selectively kept details about the task are essential for correct AI reasoning.
• Models can provide useful guidance when given a clean copy that omits sensitive PII and private notes.
• Defining a job upfront rather than a file allows the workflow to focus on actionable work while protecting privacy.
• A safe workflow includes rebuilding approved material into a new document instead of altering the original, reducing hidden metadata risk.
• The broader message is that privacy and speed are compatible when working with AI for sensitive materials.
• More videos from this channel:
  • How Do Disposable AI Agents Coordinate Over Time?
  • How to Keep Long-Running AI Agents on Track
  • How will GPT-5 reshape AI and enterprise use
• More from this channel: AI News & Strategy Daily | Nate B Jones

Questions & Answers

Q: What is the main idea behind using AI on sensitive files without uploading them?

The main idea is to separate the information the model needs to perform a task from the sensitive data that is present in the original file. By creating a sanitized copy and defining protected terms, you can get useful AI output while keeping confidential details off the AI input. This approach aims to maintain privacy without sacrificing workflow efficiency.

Q: How does Airlock assist with protecting sensitive terms in documents?

Airlock allows you to define protected terms such as customer names, project codes, or internal identifiers so the AI can recognize patterns and context without exposing private data. It uses these inputs to determine what to keep and what to hide, ensuring that only non sensitive context is shared with the model. This reduces data leakage while preserving task relevance.

Q: What does the video suggest about rebuilding the file instead of redacting it?

The video suggests rebuilding the approved material into a separate Word document rather than applying redactions in place. This avoids leaving behind hidden metadata, comments, or track changes that could reveal sensitive information. A fresh copy provides a clean, reviewable input for AI work while protecting privacy.

Q: What kind of information should be kept in the model input according to the talk?

Information that is essential for the task, such as operational assumptions, timing, and relationships relevant to the plan, should be kept. Personal details like home addresses, API keys, and private medical notes should be removed. The aim is to preserve necessary business context while excluding sensitive identifiers.

Q: What is meant by the term protected terms in Airlock, and why is it important?

Protected terms are confidential phrases or identifiers flagged to be treated with extra privacy safeguards. They are important because they enable the system to understand the task context without exposing sensitive content. This approach helps maintain accuracy in the AI output while preventing leakage of confidential information.

Q: Why is the common advice not to upload sensitive data considered insufficient in this talk?

Because in real work you need substantial context to get useful AI results, and much of that context is embedded in files. Simply avoiding upload is not practical when you want speed and accuracy. The approach shown demonstrates how to keep data private while still leveraging AI by selecting what the model actually needs.

Q: What practical steps are recommended to ensure safe AI use on confidential work products?

Create a sanitized copy of the document that omits personal data, define protected terms to preserve essential context, and use a separate rebuilt file for AI analysis. Review the sanitized copy before feeding it to AI, ensure no sensitive identifiers are present, and only include content necessary to perform the task. This reduces risk while enabling productive AI use.

Q: How does this approach address the risk of data leakage through frontier models?

By removing sensitive content before giving the document to any frontier model and only including necessary task context, the risk of leakage is reduced. The strategy also involves keeping the original file on the user’s computer and using a clean copy for AI work, which minimizes exposure while enabling efficient collaboration and iteration.

Summary & Key Takeaways

• A sanitized workflow lets you extract value from sensitive documents by separating necessary task data from confidential content while keeping the original file intact.

• The key idea is to define protected terms and rebuild a safe working document, so you can still leverage frontier models without exposing private information.

• The approach highlights that not all data needs to be visible to AI; critical context can be retained through a carefully constructed separate document.