Security review is best effort and focused on the latest release.

Older releases may not receive backported fixes unless the issue is severe and practical to address.

If GitHub private vulnerability reporting is available, please use it.

If private reporting is not available, open a minimal public issue that does not include secrets, tokens, local private paths, private logs, or model files. State that you have a security concern and provide enough high-level context for maintainers to follow up.

Do not attach logs that contain secrets.

MLX Server Manager is designed for local use. The recommended host is 127.0.0.1.

Do not expose mlx_lm.server publicly without understanding the network and security implications. OpenAI-compatible endpoints may be callable by any client that can reach the host and port.

Current release assets are unsigned local-use macOS app builds and are not notarized.

macOS Gatekeeper may warn when opening the app, including warnings that the app is damaged and cannot be opened. Verify the Release asset source, zip contents, and checksum before deciding whether to remove quarantine from an extracted app.

The app should not handle or store real API secrets for local mlx_lm.server use. The default API key value is a placeholder for OpenAI-compatible client configuration.

Do not commit:

• .env
• HF_TOKEN
• runtime settings
• model profiles
• logs containing secrets
• model files
• Hugging Face cache

Do not commit:

• settings.json
• models.json
• .app
• .zip
• .dSYM
• DerivedData
• build artifacts
• model files
• model directories
• logs
• .venv

Stop and Restart should only affect processes started and held by MLX Server Manager.

The app must not stop external mlx_lm.server processes.

The app must not rely on broad process-name matching for termination.

MLX Server Manager is not in the inference request path.

OpenAI-compatible client -> mlx_lm.server -> MLX model

The app should not become an inference proxy, Chat UI, or multi-backend wrapper.