rag-observability-evals
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The groundedness evaluation logic in
groundedness.pyinterpolates untrusted data (retrieved contexts and generated answers) into an LLM prompt. This creates a potential surface for indirect prompt injection where malicious content in documents could attempt to manipulate the judge's score. - Evidence:
- Ingestion points: The
answerandcontextsvariables ingroundedness.pyare sourced from RAG pipeline outputs. - Boundary markers: The skill utilizes a system prompt instructing the model to treat data as untrusted quoted data and ignore instructions within it.
- Capability inventory: The evaluator has the capability to perform LLM completions via the OpenAI API and write result summaries to local files.
- Sanitization: The implementation uses OpenAI's
json_objectresponse format to ensure structured output and minimize hijacking of the output stream.
Audit Metadata