Please do not open public issues for security reports.

Contact: security@yesterday-ai.de (PGP optional).

We'll acknowledge within 72 hours and provide a timeline for fixes.

• The browser-side inference runtime (MediaPipe, Gemma model loading, OPFS caching)
• The chat persistence layer (Dexie/IndexedDB handling)
• PWA service worker + cache strategy
• Any build-time tooling (Vite plugins, copy-wasm script)

• Vulnerabilities in upstream dependencies (@mediapipe/tasks-genai, @assistant-ui/react, etc.) — report to those projects directly
• Model-content safety (jailbreak prompts, harmful generations) — Gemma's license + Google policies apply
• Browser-zero-days

• No server, no API keys in client code. Nothing to exfiltrate server-side.
• Models are public. litert-community/gemma-4-* are Apache-2.0, no gated secrets.
• User data stays local. Conversations live in IndexedDB, models in OPFS — never transmitted.
• Third-party fetches: Only huggingface.co (model download) and the app's origin. No analytics, no telemetry.

• CI runs Gitleaks on every PR to catch leaked credentials
• Strict CSP + COOP/COEP headers enforced for SharedArrayBuffer
• No eval, no dynamic import() of remote code outside the MediaPipe WASM loader