learn-workbuddy is an educational clean-room harness, not a production sandbox. The code is designed to teach where safety boundaries belong: permission gates, path checks, audit trails, tool-result externalization, and explicit provider configuration.

Security reports are welcome for:

• accidental secret disclosure in the repository
• unsafe default commands in teaching demos
• path traversal or workspace escape bugs in mini_workbuddy
• misleading documentation that could cause readers to trust the teaching harness as a production sandbox
• clean-room boundary violations

This project does not accept reports that require proprietary WorkBuddy source code, private package contents, private prompts, user data, or bypass details for third-party products.

Open a private maintainer contact channel if available on the hosting platform, or create a GitHub issue with a minimal public description and no sensitive payload. Do not paste API keys, local logs, package excerpts, private prompts, or machine-specific evidence into public issues.

Never commit real credentials. Use .env.example as the template and keep real values in .env, which is ignored by git. Before publishing a branch, run:

python3 -m pytest -q
python3 scripts/verify.py

The project-specific scanner is useful, but it is not a replacement for a general-purpose secret scanner such as gitleaks or trufflehog.

The mini harness intentionally uses simple teaching policies. A real desktop agent needs stronger boundaries: OS-level sandboxing, filtered subprocess environments, network egress controls, signed extension distribution, and human approval for high-risk actions.

See docs/security-boundaries.md for the detailed safety model.