The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The documentation instructs users to execute npx shadcn@latest add https://ui.inference.sh/r/tools.json. This pattern downloads and installs code from an external URL that is not on the Trusted External Sources list.
COMMAND_EXECUTION (LOW): The skill relies on npx commands for installation and component management, which is a standard but noteworthy execution pattern for UI libraries.
INDIRECT PROMPT INJECTION (LOW): The skill serves as a UI layer that processes and displays untrusted data.
Ingestion points: args prop in ToolCall and result prop in ToolResult components.
Boundary markers: None explicitly implemented in the provided code snippets to distinguish between data and instructions.
Capability inventory: The components facilitate tool execution (onApprove={() => executeTool()}) and human-in-the-loop flows.
Sanitization: No evidence of sanitization or escaping of the displayed tool outputs is present in the snippets.
Mitigation: The inclusion of ToolApproval components for human-in-the-loop verification is a strong security best practice that mitigates the risk of automated obedience to injected instructions.

tools-ui