llm-app-patterns
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documentation provides prompts and code snippets for RAG and Agent architectures that interpolate untrusted external data (e.g., retrieved document context, user queries) into LLM prompts.
- Ingestion points: The
RAG_PROMPT_TEMPLATE,REACT_PROMPT, andPromptTemplateclass inSKILL.mdaccept external variables like{context},{question}, and{text}. - Boundary markers: The prompts use simple text headers (e.g., 'Context:', 'Question:') which are weak delimiters against adversarial input attempting to escape the intended context.
- Capability inventory: The patterns describe agents with tool-calling capabilities (
ReActAgent,FunctionCallingAgent) and access to external vector databases. - Sanitization: The code examples do not include input sanitization, filtering, or escaping logic for the interpolated data, relying on the user to implement these safeguards.
Audit Metadata