The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection by teaching the agent how to ingest and process data from external, potentially untrusted sources.
Ingestion points: The skill uses Web.Contents, Sql.Database, SharePoint.Files, and Folder.Files to pull data from the web, databases, and file systems in both SKILL.md and references/m-patterns-cookbook.md.
Boundary markers: The provided M code templates do not implement specific delimiters or instructions for the AI to ignore instructions potentially embedded within the ingested data.
Capability inventory: The skill demonstrates capabilities for network access (Web.Contents), database interaction (Sql.Database), and file reading (Folder.Files, Csv.Document, Excel.Workbook).
Sanitization: While the skill includes data cleaning examples (e.g., Text.Trim, Text.Lower), it does not provide specific sanitization to prevent adversarial data from influencing logic.
[CREDENTIALS_UNSAFE]: The skill provides patterns for handling authentication headers that follow security best practices.
Evidence: In references/m-patterns-cookbook.md, the example for handling authentication headers uses a variable (ApiKey) rather than hardcoding secrets, which aligns with safe secret management practices.
[DATA_EXFILTRATION]: The skill contains patterns for network operations and database connections, which are essential for its stated purpose.
Evidence: Functions such as Web.Contents and Sql.Database are used to fetch data from external endpoints. These are standard ETL operations and are not used for unauthorized data transmission in the provided examples.

power-query-m