patent-reader

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill interprets external patent documents and web search results, which presents an indirect prompt injection surface.\n
  • Ingestion points: Patent PDF text via tools/extract/extract_patent_text.py and web content from CNIPA results via tools/crawl/cnipa_epub_crawler.py.\n
  • Boundary markers: Structured markdown templates in assets/patent_note_template.md and explicit instructions in prompts/patent_plain_reader.md to focus on verifiable factual extraction.\n
  • Capability inventory: File system write access to the workspace and user's Obsidian vault; execution of specific configuration commands (setx, obsidian CLI) via subprocess.run.\n
  • Sanitization: Content is sanitized using HTML stripping logic (_strip_html) and summary normalization in tools/vault/clue_vault.py.\n- [COMMAND_EXECUTION]: Local utility scripts use subprocess.run to manage the user's environment, such as setting environment variables on Windows using the setx command or updating metadata in Obsidian using its official command-line interface. These operations are restricted to configuration and vault management tasks.\n- [EXTERNAL_DOWNLOADS]: The tool programmatically retrieves patent assets from Google Patents and the official CNIPA gazette portal. These sources are official or well-known and are essential for the skill's primary purpose.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 08:54 AM
Security Audit — agent-trust-hub — patent-reader