agentsop-reranker-stage
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 RAG enhancement pattern where external document chunks are retrieved and then processed by an LLM, which introduces an inherent attack surface for indirect prompt injection if the source documents contain malicious instructions.
- Ingestion points: External document data entering the pipeline context via bi-encoder and cross-encoder retrieval stages.
- Capability inventory: The instructions facilitate the creation of code that performs network API calls (Cohere, Voyage) and local file system operations.
- Boundary markers: The SOP recommends truncating the candidate pool to a narrow tail (top 3-5 chunks), which limits but does not eliminate the injection surface; it does not explicitly mandate the use of instruction-isolation delimiters.
- Sanitization: The skill does not describe specific methods for sanitizing, escaping, or filtering external content before interpolation into prompts.
Audit Metadata