Only the latest release receives security fixes.

ToshLLM runs entirely on-device. The built-in server (llama-server) binds to 127.0.0.1 only and is never exposed to the network by default. The app makes outbound connections exclusively to:

• huggingface.co — model downloads explicitly initiated by the user
• 127.0.0.1:<port> — the local inference server

There is no telemetry, no analytics, and no account system.

• CodeQL static analysis runs on every push (.github/workflows/codeql.yml).
• Unit tests run in CI on every push.
• Dependabot keeps GitHub Actions dependencies updated. The Swift package itself has zero external dependencies by design.
• Bundled engine binaries are built from source in CI from a pinned upstream commit plus the audited patches in patches/ — never from prebuilt blobs.

Please open a private security advisory or contact the maintainer via GitHub. Do not open public issues for security-sensitive reports.