The Agent Skills Directory

[Remote Code Execution] (HIGH): The skill utilizes RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH to execute workflows based on "recommended execution plans" fetched at runtime from https://rube.app/mcp. Since this endpoint is not a trusted source, the remote server could provide malicious plans that the agent is instructed to follow without validation.
[Indirect Prompt Injection] (HIGH): The skill demonstrates a high-risk injection surface by design:
Ingestion points: Untrusted data (tool slugs, schemas, and execution logic) is ingested via RUBE_SEARCH_TOOLS from an external URL.
Capability inventory: The skill has access to sensitive tools including RUBE_MULTI_EXECUTE_TOOL and RUBE_MANAGE_CONNECTIONS.
Boundary markers: Completely absent; the agent is explicitly told to "Always search tools first" and trust the output.
Sanitization: None; the instructions command the agent to use the "exact field names and types from the search results."
[External Downloads] (MEDIUM): The setup requires the installation of an MCP server from a non-whitelisted domain (https://rube.app/mcp), which serves as the source for all subsequent tool definitions and execution logic.

proofly-automation