Saturday, 21 November 2026 Press Releases Newsletter RSS
Breaking
WhatsApp Business Account Hacked in Pakistan: Full Recovery Guide Special Technology Zones Authority (STZA) & IT Park Expansion in Pakistan: Islamabad, Karachi, and Lahore Ground Report Pakistan Software Export Board & Industry Consortium Launch 'TechBridge Pakistan 2026' to Connect 500 Startups With Global Markets Global Tech Safety Summit 2026: International Standards for Frontier AI and Data Security Announced WhatsApp Business Account Hacked in Pakistan: Full Recovery Guide Special Technology Zones Authority (STZA) & IT Park Expansion in Pakistan: Islamabad, Karachi, and Lahore Ground Report Pakistan Software Export Board & Industry Consortium Launch 'TechBridge Pakistan 2026' to Connect 500 Startups With Global Markets Global Tech Safety Summit 2026: International Standards for Frontier AI and Data Security Announced
WritePures

You now have models running and an API serving them. This chapter is the one that keeps that from becoming a problem.

The mistake people actually make

Binding the API to 0.0.0.0 and exposing it to the internet.

Ollama has no authentication. None. If the port is reachable, anyone can send prompts, consume your hardware, and — if your application passes context into prompts — potentially read it.

Researchers have repeatedly found thousands of unauthenticated AI endpoints exposed online, including the tooling used in real intrusions this month. If you need remote access:

  • Put it behind a reverse proxy with authentication — Nginx or Caddy with basic auth is a ten-minute job.
  • Or use a VPN or SSH tunnel and keep it bound to localhost:
    ssh -L 11434:localhost:11434 user@your-server
    Then it is reachable on your laptop and nowhere else.
  • Firewall the port regardless, as a second layer.

"Open weight" is not "open source"

This distinction is routinely blurred and it matters commercially.

Open source traditionally means you can see and modify everything, and use it for any purpose. Most of these models are open weight: you get the finished file and permission to run it, under a licence with conditions.

Conditions that appear in real model licences include:

  • Restrictions on commercial use above a user threshold.
  • Prohibitions on using outputs to train competing models.
  • Attribution and naming requirements.
  • Acceptable-use policies covering specific applications.

Licences also differ sharply between families, and between versions of the same family. Read the licence for the specific model and version you intend to ship. "It was free to download" is not a defence, and this is the single most common way a local-model project creates a legal problem for its owner.

What stays true about privacy

The good news is genuinely good: a locally running model makes no network request to its publisher. Your prompts are not logged by anyone, not used for training, and do not leave the machine.

Two qualifications. First, that applies to the model — not to whatever you build around it. An application that logs prompts to a hosted service has reintroduced the problem. Second, anyone with access to the machine can read those logs, so treat them like any other sensitive data.

The limits, stated honestly

  • Knowledge cutoffs. A local model knows nothing after its training data ends, and has no way to look anything up unless you build that.
  • They are confidently wrong. Smaller models hallucinate more than frontier ones, and local models do not have a web search to catch themselves with. Verify anything factual.
  • Context limits. Smaller than hosted models, and filling the window costs memory and speed.
  • No updates. Permanence cuts both ways: the model that cannot be taken away also cannot be improved. Pulling a newer version is your job.

Where to go next

You can now install a runtime, choose a model that fits your hardware, run four different families, pick a sensible quantisation, call it from your own code and avoid the obvious ways to get hurt.

Reasonable next steps: build retrieval over your own documents so the model can answer from them; try a coder-specific model in your editor; or run a model from Hugging Face that is not in Ollama's library, which is where the GGUF knowledge from chapter eight earns its place.

And the habit worth keeping from chapter one: if a model becomes part of how you work, keep your own copy. That is the whole advantage of weights nobody can recall — but only if you actually hold them.