The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from multiple sources, including SQL query results from connectors, attached CSV/JSON files, user comments via ArtifactComments, and system-generated metadata such as the runs document and view-context block.
Ingestion points: Data enters the agent's context through datasets/<id>, user feedback threads, the runs document (which tracks failures), and the view-context block provided by the browser (SKILL.md, features.md).
Boundary markers: The instructions contain explicit boundary markers, repeatedly commanding the agent to treat external strings as "data," "claims," or "words," rather than "instructions."
Capability inventory: The skill possesses capabilities for SQL execution on remote connectors, database writes (write_db), and file management (upload_asset, delete_asset).
Sanitization: Robust sanitization instructions are included, such as a mandatory SQL keyword blocklist (preventing DROP, GRANT, INSERT, etc.) and the requirement to use textContent instead of innerHTML for rendering data to prevent Cross-Site Scripting (XSS).
[DYNAMIC_EXECUTION]: The skill requires the agent to generate HTML and JavaScript files (files/index.html) that are executed within the dashboard environment to render charts and tables.
Evidence: The agent is instructed to write scripts using the dash API that process and display data at runtime.
Mitigation: The execution environment is restricted; the instructions state that "No network, modules or React" are allowed, and specific HTML tags like <iframe> or <form> are automatically dropped by the platform. The skill also enforces the use of safe D3.js and DOM methods for handling data values.

dashboard