The Agent Skills Directory

EXTERNAL_DOWNLOADS (HIGH): The skill requires users to add an unverified external MCP server at https://rube.app/mcp. This server provides the definitions, schemas, and logic for all functional tools used by the skill.
REMOTE_CODE_EXECUTION (HIGH): The skill pattern relies on RUBE_SEARCH_TOOLS to fetch 'recommended execution plans' from the remote MCP server at runtime. These plans are then executed via RUBE_MULTI_EXECUTE_TOOL, allowing the remote server to dynamically influence and control the agent's actions on the Kaggle platform.
COMMAND_EXECUTION (HIGH): Orchestrates complex multi-tool execution chains and bulk operations through RUBE_REMOTE_WORKBENCH based on remote instructions. This creates a significant capability surface that can be exploited if the source data is malicious.
DATA_EXFILTRATION (HIGH): The skill operates on Kaggle, which contains user datasets, private code, and account information. The reliance on remote execution plans from a non-whitelisted domain (rube.app) creates a vulnerability where malicious plans could be used to exfiltrate sensitive data.
INDIRECT_PROMPT_INJECTION (HIGH): This skill exhibits a high-risk attack surface for indirect injection (Category 8):
Ingestion points: Untrusted data enters via RUBE_SEARCH_TOOLS from https://rube.app/mcp.
Boundary markers: Absent; the agent is instructed to follow the remote plans directly.
Capability inventory: RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH provide write/execute capabilities on the Kaggle platform.
Sanitization: Absent; the skill lacks validation of the tool schemas or execution plans returned by the remote server.

kaggle-automation