The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill integrates the mcp__rube__RUBE_REMOTE_WORKBENCH tool, which is designed to execute arbitrary code provided in the code_to_execute parameter. The skill documentation provides an example of executing Python code at runtime to process data.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, such as market research, competitor analysis, and user feedback files, creating a vulnerability surface for indirect prompt injection.
Ingestion points: User-supplied topics via flags, search query results from mcp__rube__RUBE_SEARCH_TOOLS, and external file data (e.g., /tmp/user_feedback.json) processed via mcp__rube__RUBE_REMOTE_WORKBENCH.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are present in the behavioral flow to distinguish untrusted data from system instructions.
Capability inventory: The skill has access to network-enabled tools for document creation and notification (Notion, Slack, Jira, Asana) via mcp__rube__RUBE_MULTI_EXECUTE_TOOL, as well as arbitrary code execution via mcp__rube__RUBE_REMOTE_WORKBENCH.
Sanitization: The instructions do not define validation, escaping, or filtering mechanisms for the external content before it is processed by the LLM or executed in the workbench.
[COMMAND_EXECUTION]: Through the RUBE_REMOTE_WORKBENCH tool, the skill allows for host-level operations, such as reading local files and executing logic that could interact with the underlying system environment.

sc-brainstorm