The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from remote MCP servers, which serves as a potential vector for indirect prompt injection if those servers provide untrusted or malicious output.
Ingestion points: Data retrieved via tools/call, prompts/get, and resources/read is processed in SKILL.md.
Boundary markers: The helper function does not implement specific delimiters or 'ignore' instructions for the retrieved content.
Capability inventory: The skill uses the fetch API for network communication and implies the ability to cache data to the file system.
Sanitization: No explicit content sanitization or validation logic is provided beyond standard JSON parsing.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill performs network operations using fetch to a remote server. This behavior is the primary and intended function of an MCP client.
[CREDENTIALS_SAFE]: The instructions contain explicit warnings against hardcoding or manually managing authentication tokens, recommending reliance on the platform's native connector authentication injection.

mcp-client