Self-hosting keeps the important content on hardware you control: files, photos, messages, and AI prompts. But some network traffic still leaves, including DNS lookups, certificate checks, time sync, app updates, push notifications, and federation metadata. This guide explains what stays local, what still leaks, and how to reduce it by changing the architecture.

77% of workers paste company information into AI tools, often through personal accounts with no enterprise protections. This post explains why the behavior is structural, not careless, why bans and policies fail, and why running AI locally is the only architectural fix that keeps sensitive data from leaving your control.

As of a policy that takes effect August 17, 2026, Atlassian defaults Jira and Confluence cloud customers into using their content to train Atlassian's AI products. Two data types are involved: de-identified metadata and the in-app content your team writes. Every plan can opt out of the in-app content in about five minutes (Admin, Security, Data contribution). Most plans below Enterprise cannot opt out of the metadata. Enterprise has both off by default.