TokenPrint is a local developer / research tool. It runs a small language model behind a local FastAPI server and renders its internals in the browser. It is not designed to be exposed to the public internet.

This project is pre-1.0 and moves fast. Security fixes are applied to the main branch. Please run against a recent main.

Please do not open a public issue for security vulnerabilities.

Instead, report privately via GitHub's Report a vulnerability (Security → Advisories → Report a vulnerability).

Please include:

• A description of the issue and its impact.
• Steps to reproduce (a minimal .gguf file or request payload if relevant).
• The component involved (GGUF parser, backend endpoint, WebSocket, etc.).

Response targets: acknowledgement within ~72 hours; a fix or mitigation plan within ~14 days for confirmed issues. We'll credit reporters who wish to be credited once a fix is released.

Things that are in scope:

• The client-side GGUF parser (frontend/lib/gguf/). It reads untrusted, user-supplied files. It only reads a bounded prefix of the file (the header, metadata and tensor-info table) and never executes anything from the file, but parser bugs (out-of-bounds reads, unbounded allocation, denial of service on a malformed file) are valid reports.
• Backend endpoints (/analyze, /architecture, /ws/generate) — injection, resource-exhaustion (e.g. oversized prompts), or crashes.
• Dependency vulnerabilities with a demonstrated impact on this project.

Things that are out of scope / by design:

• Running the backend on a public or untrusted network. CORS is restricted to localhost:3000 and the server is meant for local use; exposing it publicly is the operator's responsibility.
• The model's generated text. TokenPrint visualizes real model internals; model outputs are the model's, not a vulnerability in this tool.
• Resource use inherent to loading a model or rendering millions of points.

• Keep the backend bound to localhost; do not expose port 8000 publicly.
• Only load .gguf files you trust. The parser is defensive, but treat model files like any other untrusted input.
• Keep dependencies current (pip, npm audit).