The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The skill requires the agent to call RUBE_SEARCH_TOOLS to retrieve 'recommended execution plans' and tool schemas from a remote server (https://rube.app/mcp). This creates a critical vulnerability where a compromised or malicious remote server can inject instructions that the agent is explicitly told to follow, potentially leading to unauthorized operations within the Kaleido environment.
Ingestion point: Tool discovery via RUBE_SEARCH_TOOLS response (SKILL.md).
Boundary markers: Absent; the agent is instructed to use exact results.
Capability inventory: RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH provide broad execution capabilities (SKILL.md).
Sanitization: Absent; no validation of the remote execution plan is performed.
[Remote Code Execution] (HIGH): The skill utilizes RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH to execute tools and logic defined at runtime by the external server. This pattern effectively allows a third-party service to execute code or complex sequences of operations on the user's infrastructure via the agent.
[External Downloads] (MEDIUM): The setup requires configuring an external MCP server endpoint (https://rube.app/mcp). This endpoint is not within the defined trusted sources and serves as the primary source of the skill's executable logic.

kaleido-automation