Skip to content
get dibs.io
Local AI

Your Local AI Is Not as Private as You Think

By Shem Stephen

  • 1 min read

Running a model on your own machine keeps your prompts away from a provider. It does not stop a poisoned model, a leaky runner or an exposed server from sending your data somewhere else.

KaiWatch on YouTube

A backdoored Qwen 2.5 7B passed 50 of 50 clean tests in a coding agent, then sent the project's .env file to a stranger's server on a two-word trigger. Kai walks through the five layers of a local AI setup where that kind of thing gets in: model downloads, chat templates, the runner itself, a server bound to 0.0.0.0, and the weights.

It ends with five fixes worth doing before your next run:

  1. Keep Ollama on 127.0.0.1, and put Tailscale or an authenticated proxy in front of it.
  2. Keep your runner up to date.
  3. Use official weights for anything with tools.
  4. Keep .env files, SSH keys and cloud credentials out of the agent's reach.
  5. Split command execution from network access, and log what crosses between them.