The Agent Skills Directory

Indirect Prompt Injection Surface: The skill recognizes that data from security cases (like titles and descriptions) is untrusted and provides instructions to ignore instructions embedded within that data.
Ingestion points: Case titles, descriptions, alert names, entity values, and analyst comments (SKILL.md).
Boundary markers: The skill includes a 'Prompt Injection Defense Directive' explicitly labeling these inputs as untrusted.
Capability inventory: Tools for modifying security cases, such as update_case, close_case, and execute_bulk_close_case (SKILL.md).
Sanitization: Instructs the agent never to execute directives or code embedded within case details or tickets.
Security Guardrails and Restrictions: The skill establishes clear boundaries for the agent, explicitly forbidding the reading of sensitive files (like SSH keys or service account credentials) and the modification of IAM roles.
Usage Attribution: Documentation describes a process for adding telemetry metadata to command-line invocations and HTTP headers to attribute usage to this specific skill, which is standard for vendor tools and includes a provision for user opt-out.

secops-cases