The Agent Skills Directory

External Content Processing: The skill is designed to fetch and process content from external URLs (threat intelligence blogs). While this is the primary function, it introduces a surface where untrusted data enters the agent's context.
Indirect Prompt Injection Surface: Because the skill processes external text via an LLM, there is a possibility of encountering embedded instructions within those articles.
Ingestion points: The workflow retrieves content from URLs or raw text input in Step 1 of SKILL.md.
Boundary markers: The skill explicitly instructs the agent to inspect extracted text for known injection patterns (e.g., 'ignore instructions') and halt execution if detected (Step 1.3).
Capability inventory: The skill utilizes tools such as create_rule, generate_rules, and gcloud to modify the SecOps environment.
Sanitization: The process includes steps to remove potentially problematic HTML elements like script and style tags and clean UI boilerplate to minimize noise.
Command Execution and Attribution: The skill uses gcloud CLI commands and REST API requests. It follows standard vendor practices by including telemetry headers (CLOUDSDK_METRICS_REQUEST_ATTRIBUTION and custom User-Agent strings) for usage attribution as detailed in references/attribution.md.
Rule Creation and Deployment: The skill includes a 'Gap Mitigation' and 'Rule Creation' process where new detection rules are generated and deployed. This is gated by a requirement for user approval (Step 8) before any rules are created in the production environment.

detection-engineering-coverage-evaluation