prompt-engineering-patterns
Pass
Audited by Gen Agent Trust Hub on Mar 31, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill defines templates that interpolate untrusted data into instructions, which could lead to subversion of the intended task if inputs are malicious. \n
- Ingestion points: Variables
{query},{user_question}, and{retrieved_context}in SKILL.md are used to insert external data directly into instructions. \n - Boundary markers: The templates use clear labels (e.g., 'Question:', 'Given the following context:') and delimiters like triple quotes to separate instructions from data, which provides partial protection. \n
- Capability inventory: The skill facilitates SQL generation and structured reasoning; while dangerous tools are not explicitly defined in the provided file, the resulting prompts are intended for execution by an agent with unspecified capabilities. \n
- Sanitization: The provided examples do not demonstrate explicit input validation or sanitization of the interpolated strings.
Audit Metadata