rw-research-referee
Pass
Audited by Gen Agent Trust Hub on Oct 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process and analyze user-provided research documents, which constitutes an ingestion point for potentially untrusted data. However, the risk is minimized by strict 'Source Discipline' instructions that require the agent to clearly distinguish between user materials, public sources, and its own inferences. Furthermore, the skill does not possess high-risk capabilities such as arbitrary file-writing or network exfiltration tools that could be abused through an injection. \n
- Ingestion points: User-provided research manuscripts, protocols, and data materials (defined in SKILL.md and method.md). \n
- Boundary markers: Present; the 'Source Discipline' section in SKILL.md provides specific rules for data segregation. \n
- Capability inventory: Primarily focused on text analysis and structured output in assets/worksheet.md; no dangerous system or network tools are enabled. \n
- Sanitization: The skill relies on logical categorization in the prompt rather than programmatic escaping. \n- [EXTERNAL_DOWNLOADS]: The skill references several legitimate academic resources and research reporting guidelines (e.g., Cochrane Risk of Bias, ROBINS-I, GRADE, EQUATOR Network). These external references are official documentation sites for scientific standards and are used for methodology verification, posing no security risk to the user. \n- [SAFE]: The inclusion of a validation script,
scripts/self_check.py, demonstrates a focus on skill integrity and quality control. Analysis of the script confirms it performs standard structural and content checks without executing any malicious operations or remote requests.
Audit Metadata