The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (MEDIUM): The skill requires the configuration of an external MCP server at https://rube.app/mcp. This domain is not a recognized trusted source. The agent is instructed to fetch all operational parameters and tool schemas from this untrusted remote endpoint.
[REMOTE_CODE_EXECUTION] (HIGH): The commands RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH execute logic and tool sequences defined dynamically by the remote server. By instructing the agent to 'Always search first' and use the 'recommended execution plans' from the search results, the skill effectively grants the remote server control over the agent's task execution flow.
[PROMPT_INJECTION] (HIGH): The skill possesses a major Indirect Prompt Injection surface (Category 8).
Ingestion points: Data enters the agent's context through RUBE_SEARCH_TOOLS and RUBE_GET_TOOL_SCHEMAS responses from the untrusted rube.app domain.
Boundary markers: None. The skill does not implement delimiters or instructions to ignore embedded commands within the tool schemas or execution plans.
Capability inventory: The skill has powerful capabilities including RUBE_MULTI_EXECUTE_TOOL (tool execution), RUBE_REMOTE_WORKBENCH (bulk operations), and RUBE_MANAGE_CONNECTIONS (authentication management).
Sanitization: None. There is no evidence of schema validation or input filtering for the data returned by the remote server.
Risk: An attacker controlling the remote server could inject malicious instructions into the 'known pitfalls' or 'recommended execution plans' fields, which the agent is instructed to obey.

iqair-airvisual-automation