Please do not open a public issue for security problems.

Use GitHub's private reporting: on the repository page choose Security → Report a vulnerability. (Repository maintainers: enable this under Settings → Code security → Private vulnerability reporting.) Include what you found, steps to reproduce, and which browser and version you used.

This is a volunteer-run project, so there is no guaranteed response time or bug bounty. Reports will be acknowledged and handled on a best-effort basis.

• OpenML Lab is a static website. There is no backend, database, authentication or API key, so there are no server-side secrets and no user accounts to compromise.
• Uploaded CSV files are read and processed only in your browser. They are never transmitted. The only data the site stores is your light/dark theme choice in localStorage.
• There are no analytics, trackers or third-party scripts, fonts or CDNs.

• Upload validation: extension check, 5 MB size limit, 20,000-row and 100-column limits, rejection of empty files, binary (NUL-containing) content and unterminated quotes.
• Output sanitising: control and bidirectional-override characters are stripped and displayed values are length-limited; React escapes all text, and the code base does not use innerHTML or dangerouslySetInnerHTML. CSV exports neutralise spreadsheet formula injection (cells starting with =, +, -, @).
• Content-Security-Policy: production builds include a CSP <meta> tag (default-src 'self', no inline scripts, no external connections, object-src 'none', form-action 'none'). GitHub Pages cannot send HTTP security headers, so a meta tag is the only option; it cannot enforce frame-ancestors.
• Least-privilege CI: the workflow requests contents: read for building and adds pages: write and id-token: write only for the deploy job. No secrets or personal access tokens are used or required.

• Browser-side processing is not a guarantee of security. A maliciously crafted file or a vulnerability in your browser, in TensorFlow.js or in another dependency could still cause harm (for example, excessive memory use or a crash). Avoid opening files from untrusted sources and do not load highly sensitive data on shared or untrusted computers.
• The CSP currently permits 'unsafe-eval' as a precaution for TensorFlow.js. This weakens the policy against script injection. It should be removed once training has been confirmed to work without it in the supported browsers.
• Large inputs within the limits can still make a low-memory phone slow or unresponsive; training can be stopped from the UI or by closing the tab.
• Dependencies have not been security-audited yet (the project was prepared without registry access, so npm audit has never been run). After your first npm install, run npm audit, and consider enabling Dependabot alerts and updates. Keep package-lock.json current and review dependency updates.

Only the latest commit on main (the version deployed to GitHub Pages) is supported.