Qelvynox is distributed as readable source: there is no build step, no bundler and no third-party dependency, so the code you install is the code in this repository.

Open a private security advisory on GitHub rather than a public issue, so a fix can land before the problem is discoverable by anyone reading the issue list.

Please include what you found, which surface it affects and how to reproduce it. Expect a reply within a week and a fix or a written explanation shortly after.

• No telemetry, analytics or crash reporting. There is no code path that sends anything anywhere except OpenAI's own hosts.
• Session reuse. Conversation requests may reuse ChatGPT request headers and browser credentials for OpenAI services. The extension does not intentionally persist or log credentials.
• No remote code. Every script ships inside the package; nothing is fetched and evaluated at runtime. This is also why no chrome.scripting or dynamic-import permission is requested.
• No broad host access. Content scripts are limited to chatgpt.com and chat.openai.com; the one extra host permission (*.oaiusercontent.com) is used only to read image bytes, over HTTPS, and the service worker rejects the request unless it came from this extension's own scripts.
• No DOM injection into chat content. Search highlights are painted with the Custom Highlight API, and the speed booster hides wrappers rather than removing nodes React owns.
• Escaped exports. Every message string written into an exported HTML, DOCX or PDF document is escaped first, and image data URLs are validated against the base64 alphabet before they are embedded.

• Keep the data-qelvynox ownership tag on every injected node: the observer's self-mutation filter depends on it, and without it the extension re-triggers itself in a permanent loop.
• Any new innerHTML must take only literals plus esc()-ed values; user content goes through textContent.
• Never add a permission without a matching justification in docs/store-listing.md.