This policy covers the drowse.ai site: the hosted shell, the browser runtime, artifact verification, and the /api/contact Worker. For the Python package, drowse serve, and model or checkpoint loading, see the drowse security policy.

Please report security issues privately rather than filing a public issue.

• GitHub: use private security advisories
• Email: contact@drowse.ai

Please include a description, reproduction steps, the page or build revision (the drowse-source-revision meta tag), the browser and GPU, and the model involved. Do not include private prompts or credentials in the report.

Only the build currently deployed at drowse.ai receives fixes.

Hosted conversations, fitted artifacts, and model downloads are stored locally in browser storage. They are not encrypted by Drowse. The service worker caches application assets; it does not cache conversation API responses. Published SAE descriptions for the verified Gemma Scope 2 packs ship as application assets. Missing descriptions fall back to Neuronpedia, sending only the public dictionary/feature identity without cookies, referrers, activation values, or conversation text.

"Clear all local data" stops persistence and clears the application's saved sessions, artifacts, and preferences, including legacy Polythetic/Saklas databases and files that migration could otherwise restore on reload. Legacy directory handles remain usable after their contents are removed. A failed deletion is reported and can be retried; it must not silently report success. Browser profiles, extensions, OS accounts, and external backups remain outside this deletion boundary.