Please do not open a public issue for a security problem. Use GitHub's private "Report a vulnerability" button under the repository's Security tab to open a private advisory. Include steps to reproduce and the impact. You'll get a response as soon as possible.

learn-everything is local-first and privacy-conscious by design:

• Your learning data stays on your machine. Everything under tracks/ (notes, cards, the memory digest, question logs) plus profile.md, registry.json, .obsidian/, and .claudian/ are gitignored and never committed.
• Source text is sent to the host model. When you ask the tutor to learn a source, that text goes to the host model (e.g. Claude). Do not ingest material you aren't authorized to send to a third-party model — confirm first for privileged, client, or otherwise confidential documents.
• Untrusted input is treated as data, not instructions. Imperative phrasing embedded in a document or web page (including hidden/zero-width text) is flagged, never obeyed.
• Don't commit secrets. The Claudian auth token lives in its plugin settings (gitignored); never paste tokens or credentials into tracked files.

Before committing, run:

python3 scripts/security_check.py --history

The check scans tracked files, unignored new files, known ignored local runtime configs such as .claudian/ and .obsidian/, the required privacy-ignore invariants, and (with --history) reachable git history. It reports only file paths and rule names; secret values are never printed. If it flags a real token, rotate that token, remove it from local files/history, then rerun the check.

This is alpha software with no warranty (MIT). The engine is offline and dependency-free; the network surface is whatever your host model / Obsidian plugins introduce — review their security posture too.