Local does not automatically mean private.
Interfaces, logs, networks and document access. Four things to check before your first sensitive prompt.
The model is only part of the journey
Local inference can reduce sending inputs to a remote provider. But a complete application also includes a chat interface, extensions, search and possibly cloud services. Check actual data flows rather than the word “local” in a name.
Before adding a private document, verify the model, provider configuration and network features. Disconnecting the internet is a useful functional test, not proof of complete security.
Keep the API on localhost first
For personal use, keep the service available only on your computer. Do not bind to all interfaces or forward a router port without a clear reason and appropriate protection.
According to its documentation, Ollama listens on 127.0.0.1:11434 by default. Changing OLLAMA_HOST can widen access. A local API is not a substitute for authentication on a public network.
Check history and logs
Prompts and outputs may remain in chat history, logs, backups or synchronized folders. Check where the application stores data, who can access it and how it can be deleted.
Disk encryption and a locked user account address a different part of the problem than local inference. For sensitive data, consider them together with your organization’s rules.
Extensions have their own permissions
A tool that reads files or executes commands can have much greater impact than text generation alone. Grant only the access required for its task. Treat neither document text nor model output as a trusted command.
Before sharing a service with other people, separately address authentication, encrypted transport, updates and access restrictions. A local experiment and a shared service have different requirements.