Please do not open a public GitHub issue for security vulnerabilities.

Report vulnerabilities privately via GitHub Security Advisories. This keeps the details confidential until a fix is ready.

We will acknowledge your report within 48 hours and aim to release a fix within 14 days for critical issues, depending on severity and complexity.

Thuki runs entirely on your local machine: no server, no cloud backend, no telemetry. Inference happens on-device through the bundled engine (or your own Ollama install). The only outbound network requests are ones you initiate: downloading a model from the Hugging Face Hub, and the optional /search pipeline. The attack surface is limited to:

• The Tauri IPC boundary between the frontend and Rust backend
• The macOS Accessibility API integration that captures selected text and screen bounds at activation (context.rs)
• Screen capture via CoreGraphics (screenshot.rs), covering both interactive screencapture selection and full-screen CGWindowListCreateImage
• The bundled llama.cpp llama-server sidecar, which binds to 127.0.0.1 only with its web UI disabled, so nothing off your machine can reach it
• Parsing of downloaded GGUF model metadata, which is bounded and panic-safe against malformed or hostile files
• Model downloads from the Hugging Face Hub: provenance comes from pinned repo revisions, while the sha256 check is an integrity guard (truncation, bit rot, resume corruption), not a provenance control
• The local SQLite database storing conversation history
• Image processing via the image crate

We support the latest release only. Please verify you are on the latest version before reporting.