Skip to content

Latest commit

 

History

History
29 lines (22 loc) · 1.26 KB

File metadata and controls

29 lines (22 loc) · 1.26 KB

Security Model

Defaults

Task safety boundary

  • gh is the default verified backend.
  • Read first; write only when explicitly authorized.
  • Never print or persist tokens, cookies, secret values, or sensitive command output.
  • Never configure an unverified GitHub MCP.
  • Never merge, close, delete, dismiss, resolve, rerun, rollback, or force-push on inference.

Permissions

The host permission grants only explicit Task routing to the named agent. The agent denies local edit/write/task actions. Bash access is constrained by the mission and operating rules, so command choice must remain focused and read-oriented.

Threats And Responses

Threat Response
Secret exposure Redact output; never include credentials in prompts, docs, or examples.
Untrusted workflow input Inspect triggers, permissions, and untrusted data before a security decision.
False success Require command, source, GitHub, or SHA evidence and label its state.
Review authority drift Use only native review/start -> finalize -> validate.
Accidental mutation Explicit authorization plus confirmation for destructive actions.
Backend uncertainty Use gh; report MCP or permission limitations as unavailable.