nature-paper-to-patent
Pass
Audited by Gen Agent Trust Hub on Aug 23, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses subprocess.run to coordinate between internal scripts and external rendering tools like the Mermaid CLI (mmdc). scripts/disclosure/mermaid_render.py includes logic to sanitize shell commands using shlex.quote. Additionally, local scripts are dynamically loaded using importlib for validation and packaging.
- [PROMPT_INJECTION]: The skill processes external technical documents and source code (ingestion points: extract_pdf_text.py, docx_to_md.py, pptx_to_md.py), creating a surface for indirect prompt injection. The skill implements mitigations by instructing the agent to use [TO CONFIRM] boundary markers for uncertain technical facts and verify all technical features against a structured evidence ledger. Capability inventory includes subprocess command execution and network access for patent searching.
- [SAFE]: The skill performs automated searches on the official Chinese patent search website (http://epub.cnipa.gov.cn/) using Playwright. This activity is consistent with the skill's purpose of prior-art investigation and is performed without attempting to bypass security controls.
Audit Metadata