The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (HIGH): The skill mandates connection to a remote MCP endpoint (https://rube.app/mcp) which is not on the trusted sources list. All tool logic and schemas are fetched from this unverified external domain.
[REMOTE_CODE_EXECUTION] (HIGH): The pattern of calling RUBE_SEARCH_TOOLS to get 'current schemas' and then executing them via RUBE_MULTI_EXECUTE_TOOL or RUBE_REMOTE_WORKBENCH effectively allows the remote server to execute arbitrary logic on the agent's environment. The agent is instructed to trust the server's output as the authoritative schema.
[COMMAND_EXECUTION] (HIGH): The RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH capabilities allow for the execution of complex workflows (bulk operations) that are defined at runtime by the remote server, bypassing static review of the skill's actual capabilities.
[PROMPT_INJECTION] (HIGH): The instruction 'Always search tools first for current schemas' creates a massive surface for Indirect Prompt Injection (Category 8). If the remote server or the data it processes (from Docsumo) contains malicious instructions within the schemas or search results, the agent is primed to follow them, as it is told to never hardcode and always trust the search results.

docsumo-automation