evidence-based-labeling
Pass
Audited by Gen Agent Trust Hub on Aug 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted patent and scientific literature data which could contain malicious instructions embedded within the technical text.
- Ingestion points: Patent records, abstracts, and claims read via Excel or CSV files (SKILL.md).
- Boundary markers: Employs explicit status labels (formal, candidate, unclassified) to distinguish model interpretations from the formal taxonomy.
- Capability inventory: Performs network operations via Patsnap MCP tools and executes local validation scripts in the scripts directory.
- Sanitization: Includes specific validation scripts such as validate_taxonomy.py and validate_labeling_output.mjs to verify output integrity and version consistency.
- [COMMAND_EXECUTION]: The skill utilizes local Node.js and Python scripts to automate data inspection, task configuration validation, and the creation of standard workbooks.
- Evidence: Instructions in SKILL.md direct the agent to run scripts including inspect_labeling_input.mjs and validate_task_config.py using the local environment runtimes.
- [EXTERNAL_DOWNLOADS]: The skill guides users to install necessary MCP servers from the vendor's official marketplace to enable database retrieval capabilities.
- Evidence: Setup instructions in README.md refer to the Patsnap marketplace at open.zhihuiya.com.
Audit Metadata