office-review-pass
Pass
Audited by Gen Agent Trust Hub on Aug 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill processes user-provided Word documents, which creates a potential surface for indirect prompt injection via document comments or body text. \n
- Ingestion points: .docx files processed across all four operational modes (SKILL.md, scripts/review_pass.py). \n
- Boundary markers: The skill relies on behavioral instructions, directing the agent to maintain an independent perspective and engage with reviewer comments as peer input rather than commands. \n
- Capability inventory: Local file reading/writing and Word COM automation via PowerShell (references/officeops-api.md). \n
- Sanitization: No specific input sanitization for document content is detailed. \n- [COMMAND_EXECUTION]: The skill demonstrates two forms of command execution intended for legitimate operations. \n
- Subprocess usage: The test script (scripts/test_review_pass.py) uses subprocess.run to execute the local Python driver for CLI validation. \n
- COM Automation: The documentation (references/officeops-api.md) notes that the vendor package (ane_package.officeops) interacts with Word COM by driving PowerShell 5.1 scripts, which is the expected mechanism for Office automation on Windows.
Audit Metadata