The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/build_patent_package.py uses importlib.util.spec_from_file_location to dynamically load internal Python modules such as validate_patent_draft.py and audit_claims.py from calculated file system paths. This pattern can be exploited if the file system is compromised.\n- [REMOTE_CODE_EXECUTION]: The script scripts/disclosure/mermaid_render.py attempts to execute the @mermaid-js/mermaid-cli tool. If not found locally, it uses npx -y to download and run the package from the public npm registry at runtime, which introduces a supply chain risk through automatic execution of external code.\n- [COMMAND_EXECUTION]: The skill frequently uses subprocess.run to invoke external tools (like Mermaid CLI) and other Python scripts in the package. While the implementation uses argument lists (which are generally safer than shell strings), it presents a risk if user-controlled content influences the command line arguments.\n- [EXTERNAL_DOWNLOADS]: The scripts scripts/disclosure/cnipa_epub_crawler.py and scripts/disclosure/cnipa_epub_search.py perform network operations to http://epub.cnipa.gov.cn/ to scrape patent data. This interaction with a non-whitelisted domain includes logic to bypass automation detection and fetch data into the local environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes technically complex and untrusted user materials, such as research papers (PDF), source code, Word documents, and PowerPoint presentations.\n
Ingestion points: Untrusted content enters via scripts/extract_pdf_text.py, scripts/disclosure/docx_to_md.py, and scripts/disclosure/pptx_to_md.py.\n
Boundary markers: While SKILL.md requires features to be mapped to source IDs, it lacks explicit safety instructions to the agent to ignore hidden commands within the analyzed documents.\n
Capability inventory: The skill possesses capabilities for network scraping, file writing, and subprocess execution, which could be abused if an indirect injection successfully overrides agent instructions.\n
Sanitization: There is no evidence of technical content sanitization or filtering of the extracted technical prose before it is processed by the AI models.

nature-paper-to-patent