config-evals
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the setup of LLM-based evaluations that ingest and process data from external datasets and workflow outputs, which presents a surface for indirect prompt injection.
- Ingestion points: Data enters the evaluation context via
data-tables(test datasets) and workflow node outputs (specified byendNodeName). - Boundary markers: The skill uses n8n expression syntax
={{ $json.<field> }}to reference data. The instructions mandate the use of the=prefix to ensure values are treated as expressions rather than literal strings, though this does not prevent injection within the resolved content. - Capability inventory: The skill utilizes the
eval-configtool (actions: create, update, list, get, delete) anddata-tablestool (actions: list, create, insert) to manage evaluations and their associated data. - Sanitization: There is no explicit requirement or mechanism described for sanitizing or filtering the content of the workflow outputs or dataset columns before they are passed to the judge models.
Audit Metadata