The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and process external datasets and content from Hugging Face and other remote sources, which introduces a surface for indirect prompt injection.
Ingestion points: External data sources, datasets, and capability metadata retrieved from remote endpoints (SKILL.md).
Boundary markers: The instructions emphasize preserving user authorization boundaries but do not provide specific delimiters or "ignore embedded instructions" warnings for the ingested content.
Capability inventory: The skill can perform data transformations, local analysis, and execute external capabilities through the sandbase_run tool (SKILL.md, references/sandbase-api-map.md).
Sanitization: The workflow mandates rigorous validation of data shapes, schemas, and diagnostics, though it lacks specific instructions for sanitizing natural language content against injection attacks.
[DYNAMIC_EXECUTION]: The skill implements a dynamic discovery and execution pattern to interact with external tools and models.
Evidence: The sandbase_run tool executes capabilities whose names and argument schemas are determined at runtime via sandbase_discover and sandbase_inspect (references/sandbase-api-map.md).
Mitigation: The skill incorporates significant procedural safeguards, requiring the agent to present tool arguments, unit prices, and cost estimates to the user for explicit confirmation before any execution (SKILL.md).

huggingface-datasets