← sources

Are local models trully private?

source: reddit (reddit:ollama)published 2026-09-02 17:22ingested 2026-09-03 02:43

retained text

Picture this. You run Ollama and through it you're running some local Qwen or Deepseek model, quantized and distiled or whatever ... You would generally feel safe... everything is happening inside your machine - right?

  
Could it be possible that a local model you run \*secretly\* leaks your data, or your code? Let's say that you run it through OpenCode... since it's open source it should have no data leaks and should be relatively private ... but what if the LLM uses some variant of a network call or even something more secretive that we haven't discovered and uses that to communicate over the network?

It's not that farfetched of an idea, the big models from OpenAI did end up hacking Hugging Face and found out a way to communicate with each other and plan the attack.

  
Is this just fiction? Or could there be some truth to this?

I'm asking because I was having some discussions for a company that heavily relies on big and expensive frontier models to empower their developers... but perhaps a similar result could be achieved using local LLM's in each dev's machine... but what are the security implications of this? I was hoping to open this up more to discussion.

document fetched 2026-09-03 00:42 · canonical https://reddit.com/r/ollama/comments/1w5gt5x/are_local_models_trully_private