The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface as it ingests and processes untrusted metadata from external sources (Microsoft Fabric semantic models).
Ingestion points: Metadata is retrieved via tools like get_model_schema, list_models, and find_items, specifically reading measure names, descriptions, and expressions.
Boundary markers: No specific delimiters or "ignore instructions" warnings are defined for the metadata ingested from the model schema.
Capability inventory: The skill utilizes tools capable of executing queries (run_dax_query, run_sql_query) and discovering model items (list_models, get_model_schema).
Sanitization: The skill mitigates risks by mandating the use of query parameters (@name) for user-provided values in DAX and SQL queries, effectively preventing direct injection into the query syntax.
[DYNAMIC_EXECUTION]: The skill involves the dynamic construction of DAX and SQL query strings at runtime based on user prompts and discovered schema data.
Evidence: Instructions in SKILL.md and references/dax-query-patterns.md guide the agent to assemble complex EVALUATE blocks and SUMMARIZECOLUMNS calls dynamically.
Context: This behavior is central to the skill's primary purpose and follows well-defined patterns and templates, reducing the risk of arbitrary or unintended code execution.

querying-fabric