rag-evaluation
Pass
Audited by Gen Agent Trust Hub on Oct 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides template prompts for evaluating RAG outputs (e.g.,
FAITHFULNESS_PROMPT,CONTEXT_RELEVANCE_PROMPT, andUNANSWERABLE_PROMPT) that interpolate untrusted external data into the prompt of a judge model. This creates a surface where an attacker could embed malicious instructions within retrieved documents or AI-generated answers to manipulate the judge's evaluation or bypass security constraints. - Ingestion points: Untrusted data enters the judge context via
{context},{answer},{passage}, and{question}variables inSKILL.md. - Boundary markers: The prompts use basic text-based labels (e.g.,
Sources:,Answer to evaluate:,Question:) to separate untrusted data from instructions, which provides minimal protection against adversarial input. - Capability inventory: The skill uses model invocation capabilities (
client.messages.create) to process these prompts. - Sanitization: There is no evidence of escaping, schema validation, or specialized instruction-isolation techniques for the interpolated data in the provided code snippets.
Audit Metadata