The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and process content from external, live sources such as API documentation and Model Context Protocol (MCP) specification sites. This creates an attack surface where an agent could ingest malicious instructions embedded in remote content.
Ingestion points: Instructions to fetch sitemap.xml and specific documentation pages from modelcontextprotocol.io and public GitHub repositories (SKILL.md).
Boundary markers: The skill recommends defining strict input schemas using libraries like Zod (TypeScript) or Pydantic (Python) to validate and constrain tool inputs.
Capability inventory: The skill guides the creation of tools capable of performing actions on external services, such as creating GitHub issues or interacting with Slack APIs (SKILL.md).
Sanitization: The instructions explicitly warn about injection risks in UI components and advise HTML-escaping untrusted display values, such as client-controlled names, in the consent screen to prevent attacks at the platform boundary (SKILL.md).
[SAFE]: The skill references resources and documentation from established organizations and well-known technology services, including the Model Context Protocol project and official repositories on GitHub. These references are handled as safe, standard documentation lookups.

mcp-builder