guardrails-safety-filter-builder
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
classify_topicfunction creates an injection surface by interpolating user-provided queries directly into an LLM prompt without delimiters. - Ingestion points: The
classify_topicfunction inSKILL.mdaccepts thequeryvariable as an argument. - Boundary markers: The
classification_prompttemplate lacks clear delimiters (such as triple quotes or XML tags) to separate the user query from the instructions. - Capability inventory: The skill utilizes an
llm()function to process the resulting prompt, which represents the execution capability. - Sanitization: No escaping, validation, or structural filtering is performed on the
querystring prior to interpolation.
Audit Metadata