The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to send task updates and results to an external messaging platform (Lark). It mitigates risk by explicitly instructing the agent not to send secrets, tokens, or private data, and requires authorization before uploading desensitized files.
[PROMPT_INJECTION]: The skill processes messages from Lark as user input, creating an indirect prompt injection surface.
Ingestion points: External messages are ingested via lark-cli event consume as described in the process for handling user replies in SKILL.md.
Boundary markers: The skill relies on identity verification and message threading for context but does not define explicit delimiters (like XML tags) to wrap the external content within the agent's prompt context.
Capability inventory: The skill can trigger any tool or operation available to the host agent, including high-risk or destructive actions, based on the remote user's "decision" or "approval."
Sanitization: The skill performs identity verification using open_id and requires the agent to ask for clarification on ambiguous replies rather than guessing intent, reducing the risk of accidental execution.

lark-monitor