course-publisher
Pass
Audited by Gen Agent Trust Hub on Jul 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it processes external course data to generate student-facing communications.\n
- Ingestion points: The
shared/course_passport_schema.mdfile is used as the primary data source.\n - Boundary markers: Present; the skill enforces 'Iron Rules' and 'Checkpoint Protocols' that require human verification of all AI-generated drafts.\n
- Capability inventory: The skill can write files to local directories and execute specific local Python scripts for document processing.\n
- Sanitization: Absent; the skill relies on manual human review and automated consistency checks against the syllabus rather than programmatic sanitization.\n- [COMMAND_EXECUTION]: The skill references the use of local scripts
scripts/export_lms.pyandscripts/render_document.pyto automate the creation of LMS packages and document rendering. These are internal tools essential for the skill's primary function.\n- [SAFE]: The skill strictly implements a 'draft-only' protocol, explicitly stating it cannot autonomously send emails or upload to platforms like Canvas, Moodle, or Blackboard, ensuring a human remains in the loop.\n- [SAFE]: A 'fact-trace' mechanism is used to verify every date, room number, and policy against the course passport, preventing the agent from inventing or hallucinating information.
Audit Metadata