MEDIUM
The severity is rated MEDIUM because while there's a risk associated with client-side execution, the impact of an exploit would depend on the specific vulnerabilities present in Transformers.js or WebGPU. The real-world exploitability varies between homelab and production environments, but patches are likely to be developed if any critical issues arise.

The Nemotron-3-Nano (4B) is a new hybrid Mamba + Attention model from NVIDIA that runs locally in the browser using WebGPU. This technology allows for efficient computation directly on the user's hardware, potentially mitigating some server-side performance bottlenecks. However, running such models client-side introduces security risks related to code execution and data privacy. The demo utilizes Transformers.js, a JavaScript library for working with transformers, which could be susceptible to vulnerabilities in its implementation or dependencies. Engineers and sysadmins must consider the potential exposure of sensitive information when deploying client-side AI models and ensure that these libraries are up-to-date and secure.

Affected Systems
  • Transformers.js
  • WebGPU
Affected Versions: All versions
Remediation
  • Ensure that Transformers.js is updated to the latest version by running `npm install transformersjs@latest` or checking the official repository for updates.
  • Regularly review and update dependencies used in client-side applications to mitigate potential vulnerabilities.
  • Implement Content Security Policy (CSP) headers to restrict sources of content loaded by a web page, reducing the risk of code injection attacks.
Stack Impact

The impact on common homelab stacks is moderate. Developers using WebGPU for client-side computation must ensure that their JavaScript dependencies are secure and up-to-date. Specific files impacted include `package.json` for dependency management and potentially any JavaScript or HTML files where Transformers.js is implemented.

Source →