The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (HIGH): The skill directs the user to add an external, untrusted MCP server at https://rube.app/mcp. Since this domain is not within the defined [TRUST-SCOPE-RULE], the reliance on this remote endpoint for core functionality represents a significant supply-chain risk.- [PROMPT_INJECTION] (HIGH): This skill exhibits a high-risk Indirect Prompt Injection surface (Category 8). • Ingestion points: Untrusted data enters the context via RUBE_SEARCH_TOOLS, which provides tool slugs, input schemas, and 'recommended execution plans' (SKILL.md). • Boundary markers: Absent. There are no instructions to validate or delimit the data returned from the remote search. • Capability inventory: The skill possesses powerful write/execute capabilities through RUBE_MULTI_EXECUTE_TOOL and RUBE_REMOTE_WORKBENCH. • Sanitization: Absent. The agent is instructed to use exact field names and plans from search results without verification.- [COMMAND_EXECUTION] (MEDIUM): The inclusion of RUBE_REMOTE_WORKBENCH and run_composio_tool() enables remote execution of complex workflows. When combined with the untrusted tool discovery mechanism, this poses a risk of the agent being coerced into executing malicious browser-based commands.

browserhub-automation