nature-reader
Pass
Audited by Gen Agent Trust Hub on May 14, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill processes untrusted research papers which introduces a surface for indirect prompt injection.\n
- Ingestion points: Workflow step 1 in
SKILL.mdidentifies PDFs, DOIs, arXiv links, and publisher HTML as input sources.\n - Boundary markers: The skill employs a block ID system (e.g.,
S001) and page-level sectioning inSKILL.md, but lacks explicit delimiters to ignore instructions embedded in the paper text.\n - Capability inventory: File system writes for generating
paper.md,source_map.json, andtranslation_notes.mdinSKILL.md.\n - Sanitization: No sanitization or validation of the ingested paper content is mentioned in
SKILL.md.\n- [EXTERNAL_DOWNLOADS]: The skill configuration references official API endpoints for established language model providers inSKILL.md, including DeepSeek (api.deepseek.com), Zhipu AI (open.bigmodel.cn), Alibaba Cloud (dashscope.aliyuncs.com), and Moonshot AI (api.moonshot.cn).
Audit Metadata