The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill mentions a risky_eval tool in the 'Recovery' section. However, this is described in an advisory capacity, instructing the agent that scripts using this tool must manually manage failure cleanup and undo states. No malicious code is provided for execution.
[INDIRECT_PROMPT_INJECTION]: The skill describes tools like get_capabilities, get_mesh_info, and list_outline that ingest data from a running Blockbench project. While this represents a surface for indirect prompt injection if project data were maliciously crafted, the skill includes standard prompts for discovery and does not suggest executing instructions found within project data.
[COMMAND_EXECUTION]: The skill documents interactions with an external MCP server and host filesystem via tools like export_model. An example shows exporting a model to a path on a Windows filesystem (C:/models/robot.bbmodel), which is a legitimate feature for this type of agent integration.

blockbench-mcp-overview