rag-retrieval

Pass

Audited by Gen Agent Trust Hub on Apr 16, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements Retrieval-Augmented Generation (RAG) patterns that ingest data from untrusted external sources (web search results and vector databases) and interpolate it directly into LLM prompts.
  • Ingestion points: External data enters the agent context via retrieve_documents in scripts/rag-pipeline-template.ts and the WebSearcher class in scripts/scripts/crag-workflow.py.
  • Boundary markers: Prompt templates in rules/core-basic-rag.md and scripts/rag-pipeline-template.ts use simple delimiters like "Context:\n" without robust escaping or explicit instructions to the model to ignore instructions embedded within the retrieved text, creating a surface for indirect prompt injection.
  • Capability inventory: The skill is configured to use capabilities such as WebFetch, WebSearch, and database access (SQLAlchemy/PGVector), which could be leveraged by an attacker if an injection is successful.
  • Sanitization: There is no evidence of content sanitization, filtering, or validation of the retrieved data before it is presented to the language model.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 16, 2026, 10:16 AM