If you find a security issue, please do not open a public issue. Instead, email frederico [dot] kluser [at] gmail [dot] com or open a private security advisory at:

https://github.com/frederico-kluser/browser-llm/security/advisories/new

We'll respond within a week.

The relevant attack surface for an on-device LLM browser is:

• WebView sandbox escape (mostly upstream — WKWebView / Chromium).
• Malicious page exploiting our injectedJavaScript bridge.
• Maliciously-crafted GGUF triggering a llama.cpp parser vulnerability.
• Prompt injection causing the assistant to take actions the user didn't intend (current MVP has no agentic actions, so impact is limited to misleading text in the chat).

• Bugs in upstream llama.cpp, react-native-webview, or Expo. Please report those to their respective projects.
• Issues that require physical access to an unlocked device.