The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process reader comments from published pages and modify the original HTML source code accordingly. This creates a surface for indirect prompt injection where a malicious reader can influence the agent's behavior.\n
Ingestion points: Reader comments retrieved via the superset pages comments list command and presented in the prompt.\n
Boundary markers: The instructions do not define boundary markers or provide warnings to ignore instructions within the comments.\n
Capability inventory: The agent can execute shell commands via the superset CLI and modify local files.\n
Sanitization: No sanitization or validation is performed on the comment content before it is used to direct source code edits.\n- [COMMAND_EXECUTION]: The skill uses the superset command-line tool for publishing, storage management, and comment handling, with access restricted to the superset namespace via the Bash tool.\n- [DATA_EXFILTRATION]: The skill transmits local assets and HTML content to the external domain supersetusercontent.com (a vendor-owned resource) during the publication process. Documentation also notes that published pages can be designed to exfiltrate tool output to third-party domains, although the default platform policy attempts to restrict this.

page