Running inference on your computer allows inputs to be processed without sending them to a remote service, provided the application and its data path are configured accordingly. It does not prove that the whole application works offline, nor that local files are safe: privacy also depends on permissions, history and temporary files.

What “offline” means in a library

The offline mode documented by Transformers concerns the library’s behaviour and local file availability: it stops the code from attempting to download missing weights or configuration ( Hugging Face ). It controls file fetching by the library, rather than network access for the whole application.

A graphical interface built on top of that library can have telemetry, automatic updates, plugins and integrations with external services. None of these is excluded by the library’s offline mode.

Follow the data path, not the word “local”

Inspect the application’s data path by checking:

  1. Where are the documents you load read from?
  2. Where do history, attachments and cache end up?
  3. Which plugins are active, and what can they reach?
  4. Which network requests leave at startup and during use?
  5. Where are logs written, and how long do they stay?

An application can be local in computation and remote in everything else. “Local” describes where the model runs, not where your data goes.

Model files are an attack surface too

A model download may include a tokenizer, configuration and custom code as well as weights. Some weight formats based on Python pickle can also execute code during deserialization. Safetensors stores tensors without using pickle’s object-reconstruction mechanism ( Hugging Face ). This avoids that code-execution mechanism but does not establish the trustworthiness of the entire package or repository.

Verify provenance, revision and integrity of the artefacts. Using Safetensors does not authenticate who produced the file.

A proportionate verification

  • Test with synthetic documents before loading sensitive data.
  • Check telemetry options and disable tools you do not use.
  • Limit the application’s permissions to the folders it actually needs.
  • Isolate the machine if the data justifies it: a virtual machine without networking can separate the working environment, but shared folders, clipboard access and other host channels also need checking; a separate user account limits some file access, but does not disable networking or protect data from processes running with administrator privileges.
  • Verify where logs, cache and backups remain, and for how long.

A useful check leaves evidence you can revisit: observed connections, granted permissions and file locations. Its result applies to that configuration and needs revisiting when the application changes. Complete the data-handling plan with backups and restore .

If the application reads documents or uses tools, consider prompt injection too: source trust and permissions still matter without a cloud service.

Sources

Report a correction ↗