The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to install an external implementation package, flytoday-mcp, using the uvx tool. This package serves as the backend for all the tools described in the skill.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an interface for external travel data, creating a surface for potential injection through untrusted content.
Ingestion points: Tools such as ft_hotel_reviews (which fetches scores and text reviews), ft_stay (which retrieves host-provided villa details and rules), and ft_activity (which fetches activity highlights and descriptions) ingest data that is authored by third parties (guests or hosts).
Boundary markers: The instructions do not provide specific boundary markers or instructions to the agent to treat these external descriptions and reviews as untrusted data.
Capability inventory: The skill tools are explicitly defined as read-only, which mitigates the risk as the agent lacks built-in capabilities to perform write operations or network exfiltration based on malicious instructions found in the data.
Sanitization: There is no explicit mention of sanitization or filtering logic applied to the external text before it is presented to the agent.

flytoday