golden-dataset-curation
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves fetching and analyzing content from external URLs and local files to populate a golden dataset. This design creates a surface for indirect prompt injection attacks, where malicious instructions embedded in the ingested documents could attempt to override the sub-agents' evaluation logic or influence the generated metadata.
- Ingestion points: The 'Fetch Agent' in the curation pipeline retrieves content from arbitrary URLs or reads from local files as specified in the workflow architecture.
- Boundary markers: The prompt templates for the sub-agents (Quality Evaluator, Difficulty Classifier, etc.) interpolate variables such as
{content_preview}directly into the instructions. The provided examples lack explicit delimiters (e.g., XML tags or triple quotes) or negative constraints instructing the agent to ignore commands within the source text. - Capability inventory: The skill focus is on analysis and metadata generation. It leverages external tools for observability (Langfuse) and data management, but does not explicitly include high-risk capabilities like arbitrary shell execution or system-level writes.
- Sanitization: No explicit sanitization or filtering of the ingested content is described before it is passed to the LLM-based analysis agents.
Audit Metadata