KM4BI KM4BI

I removed 5 things from a file before ChatGPT saw it. It still found the problem.

Nate B Jones introduces Airlock as a way to preserve useful context for AI while stripping unnecessary sensitive data.

Nate B Jones frames a practical privacy problem: many work documents are too sensitive to upload as-is, yet they contain the context that makes AI useful. Telling people “do not paste sensitive data into AI” does not finish the job when the contract still needs review or the operating plan still needs analysis.

His demonstration uses a synthetic pricing plan and an application called Airlock. Before processing the file, the user defines protected terms such as customer names, project names, internal code names, or other phrases that only become confidential inside a specific business context.

The key idea is task-based filtering. If the model is asked to find assumptions that could break an operating plan, it needs details such as a simultaneous warehouse migration, ERP readiness before billing, and training during normal shifts. It does not need a person’s name, home address, medical note, API key, or a price that is irrelevant to the question.

Airlock rebuilds the approved material into a new Word document rather than visually redacting the original. That matters because Word files can retain comments, tracked changes, author names, old edits, and external relationships even when the visible page looks clean.

The broader point is that AI work has moved from generic chatbot prompts to real proposals, contracts, meeting notes, and codebases. The amount of context people want to provide has grown dramatically, while sensitive information remains mixed into the same files. If the safe path is slower than the upload button, users will be pushed toward shortcuts.

Nate connects this to shadow IT and security fatigue, citing Verizon telemetry showing much wider AI usage on corporate devices and frequent access through non-company accounts. His conclusion is that safety has to live in the same path as convenience: start with the job, identify the minimum useful context, and only then decide what can leave the local machine.

Source