# Security

Generated Python and model edits are untrusted. Read docs/BOUNDARIES.md before
running on a workspace. Docker is required for generated code and draft editing.
The model's provider receives the permitted history and validation feedback.

Report a suspected boundary bypass privately through the repository's GitHub
security advisory feature when available. Do not put credentials, private files,
or a live exploit against someone else's workspace in a public issue.

Only the latest alpha is maintained. The pinned upstream DSH profile may contain
other tools; maintenance sessions restrict and guard them independently. Normal
fallback sessions retain the base agent's permissions and behavior.