rag-observability-evals
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines an automated evaluation prompt (
GROUNDEDNESS_PROMPT) that interpolates external context and answer text using standard string formatting, making it susceptible to indirect prompt injection if untrusted text contains malicious instructions. - Ingestion points:
contextsandanswervariables in thescore_groundednessfunction insideSKILL.mdare passed directly into the prompt layout. - Boundary markers: The prompt uses simple text labels (
Context documents:andAnswer to evaluate:) but lacks strict structural isolation or systemic instructions to ignore embedded commands. - Capability inventory: The interpolated prompt is executed via
client.chat.completions.createusing thegpt-4omodel. A successful injection could manipulate the evaluation outcome or structural JSON compliance. - Sanitization: No input filtering, instruction escaping, or validation is performed on the data prior to formatting into the prompt template.
Audit Metadata