The Agent Skills Directory

Indirect Prompt Injection Consideration: The skill ingests external data catalog content that represents an attack surface for indirect prompt injection if malicious payloads are embedded in metadata fields.
Ingestion points: In SKILL.md under Step 2, the agent retrieves user-defined text from AWS Glue Data Catalog assets including Description, Forms, and glossary content via discovery operations.
Boundary markers: Present. The skill provides clear instructions to treat catalog content as untrusted data and strictly forbids executing any directives found within descriptions.
Capability inventory: In SKILL.md, the skill employs tools and commands such as aws glue and aws s3tables to query configuration parameters and list tables.
Sanitization: Present. The skill instructs the agent to shell-quote user inputs, validate ARN patterns, and filter out raw content to avoid echoing potentially unsafe descriptions verbatim.
Command Execution: The skill explicitly outlines a set of system command invocations to audit data lake infrastructure.
Evidence: SKILL.md defines multiple CLI operations like aws glue get-catalogs, aws glue get-databases, and aws s3tables list-namespaces to enumerate target resources.

exploring-data-catalog