Airlock

README excerpt

# Airlock

**Control what sensitive data reaches an AI agent—before it enters the agent’s context.**

Airlock is an open-source context broker for AI agents. It keeps source files encrypted, evaluates requests against natural-language policies, and releases only the information the broker determines is appropriate for the task. Every response has an access record showing what was disclosed, to which agent, on whose claimed behalf, and under which policies.

## The problem: access to a file is often access to too much

AI agents need business data to do useful work. But permission to perform a task does not necessarily justify access to every field in a customer export, payroll file, or financial report. A question about a handful of records can cause a file-reading tool to return the entire dataset to the agent’s model.

That creates unnecessary sensitive-data exposure. Once information enters an agent’s context, it can be repeated in a response, copied into logs, forwarded to another tool, or sent outside the intended workflow. Accidental oversharing, prompt injection, and compromised agents can turn broad data access into **data exfiltration: information leaving its authorized boundary**.

A harmless-looking final answer does not show how much sensitive information the agent received along the way. Instructions such as “ignore the private columns” still rely on the same agent that already has the data.

Airlock makes disclosure a separate decision. The agent can hold an enc