agentsop-map-reduce-fanout
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a framework for processing lists of independent items in parallel (map-reduce). This inherently creates an ingestion point for untrusted data (e.g., summarizing N docs, processing candidate ranks). While the skill provides architectural advice on handling these items, the underlying LLM calls are susceptible to indirect prompt injection if the input items contain malicious instructions. * Ingestion points: items list in various code snippets (SKILL.md). * Boundary markers: Generally absent in the provided prompt templates (e.g., Appendix B uses f-strings without explicit delimiters). * Capability inventory: Subprocess calls, LLM invocations, and vector database retrieval across several operation models. * Sanitization: The skill recommends human-in-the-loop or LLM-free verification checks in Step 7, but these are not enforced within the executable code templates.
Audit Metadata