nature-reader
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill workflow in 'static/core/output-contract.md' and 'static/core/workflow.md' directs the agent to run a local Python tool 'scripts/validate_reader_math.py' to ensure that the generated output meets structural requirements.
- [DYNAMIC_EXECUTION]: The skill utilizes a provided Python script to perform runtime verification of the agent's output, checking for proper math delimiters and source map consistency.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external papers (PDF, HTML, DOI), creating a vulnerability surface for instructions embedded within research materials. 1. Ingestion points: Data enters through various format-specific fragments such as 'pdf-text.md' and 'html.md'. 2. Boundary markers: The instructions mandate the use of specific Markdown block pairs ('Original'/'中文') and unique ID anchors to separate and track content. 3. Capability inventory: The agent has the capability to write local files and execute the provided validation script. 4. Sanitization: The 'validate_reader_math.py' script provides a validation layer to detect corrupted or improperly formatted data in the final artifacts.
Audit Metadata