The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides a template in references/templates/rest-endpoint.md that is vulnerable to path traversal, creating a surface for indirect prompt injection or unauthorized data exposure through unsanitized URL parameters.
Ingestion points: The params._splat variable in the splat route handler receives unvalidated path data directly from the URL.
Boundary markers: Absent. The template lacks delimiters or warnings to ignore malicious path segments (e.g., ../).
Capability inventory: The handler uses readFile to access the file system and returns the resulting content via a Response object.
Sanitization: No path normalization, validation, or escaping is performed on the user-controlled input before it is used to access the file system.
[COMMAND_EXECUTION]: The skill instructions mandate the execution of an external tool as part of the development workflow.
Evidence: The SKILL.md file contains the mandatory instruction: "After changes, run fuse-ai-pilot:sniper."
[DATA_EXFILTRATION]: The readFile template in references/templates/rest-endpoint.md facilitates the reading and transmission of server-side files to an external requester without sufficient access control or path validation.

start-server-routes