Source: How to Use AI on Files You’re Not Allowed to Upload (Video)
Host: nate-b-jones
Published: July 24, 2026
Source URL: https://www.youtube.com/watch?v=EuVvLwWZ5wc
Raw Transcript: raw/EuVvLwWZ5wc_transcript.md
Executive Summary
In this video, nate-b-jones addresses the fundamental breakdown of traditional enterprise privacy advice in the age of frontier AI. Standard enterprise guidelines rely on blanket rules like “don’t paste sensitive information into AI.” However, as AI interactions shift from empty chatbot prompts to massive, context-rich document reviews (contracts, operating plans, codebase audits), sensitive data becomes inextricably bundled with operational context.
When employees face delivery pressure without ergonomic safety tools, they suffer from security-fatigue and resort to unapproved non-company accounts (driving massive shadow IT growth). To solve this, Nate introduces airlock, a local tool he built to execute context-minimization. Instead of drawing black rectangles over complex Word documents (which retain hidden metadata and comments), Airlock allows users to define custom protected terms, strip irrelevant PII and confidential details based on task intent, and rebuild clean, isolated files for frontier model processing while leaving original files untouched locally.
Key Takeaways & Core Concepts
1. The Breakdown of Traditional Privacy Advice
- Standard corporate slides advise employees not to paste sensitive data into LLMs, but sensitive work (risk reviews, contract evaluations, operating plan audits) still must get done.
- In 2024, AI interactions involved short, empty-prompt questions. By 2026, useful AI work requires 100x to 1,000x more context from proprietary files.
- Because confidential details are bundled inside operational files, asking individual employees to act as manual privacy filters while working under tight deadlines fails.
2. Shadow IT Telemetry & Security Fatigue
- Verizon Telemetry: Enterprise data shows corporate device AI usage rose from 15% to 45%, with two-thirds of users accessing AI through non-company personal accounts. Source code was the most common material involved in data policy events.
- NIST Security Fatigue: When repetitive security decisions pile up, users defaults to the path of least resistance (the upload button).
- The Binary Enterprise Response: Responsible power-users rely on personal trust, while ultra-cautious industries (such as health care) abstain completely, leaving valuable AI leverage uncaptured.
3. Task-Intent context-minimization
- Blanket redaction (deleting every name, date, price, and role) creates a clean file that is completely useless to an LLM.
- Context Minimization starts with the job to be done, not the file:
- If auditing an operating plan for launch risks, the model needs warehouse migration timelines and ERP dependencies, but does not need customer home addresses, emails, private medical notes, or unreleased prices.
- The same data point (e.g., price) can be essential for contract negotiation but irrelevant for drafting a launch summary.
4. Rebuilding Clean Files vs. Masking Word Metadata
- Word files are complex containers containing hidden comments, track changes, author names, old edits, and external relationships that remain active even if hidden visually.
- Rather than editing or redacting original documents, airlock creates a brand new Word file containing only the approved load-bearing context. The original file remains untouched on the user’s local hardware (leveraging local-ai-safeguarding).
5. Custom Protected Terms in airlock
- Standard automated PII detectors catch social security numbers and email addresses, but miss company-specific confidential phrases (e.g., codename “Project Lantern”).
- Airlock allows users to specify custom protected terms, combining user context with automated pattern recognition under a single review screen where hiding items is the default.