The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests scripts, console output, and assets from the Roblox DataModel, creating an attack surface for indirect prompt injection.
Ingestion points: Data enters the agent via script_read, get_console_output, search_game_tree, and insert_asset (SKILL.md).
Boundary markers: The instructions do not define explicit delimiters or instructions for the agent to ignore potentially malicious content within these inputs.
Capability inventory: The agent possesses powerful tools including execute_luau (code execution), multi_edit (file modification), and http_get (network access) (SKILL.md).
Sanitization: No specific sanitization or validation logic is applied to the content retrieved from the Roblox environment.
[COMMAND_EXECUTION]: The skill setup involves executing local commands and script files to launch the MCP server (StudioMCP and mcp.bat). These commands are intended to be run by the user during initial setup.
[REMOTE_CODE_EXECUTION]: The insert_asset tool downloads third-party content from the Roblox Creator Store which can contain executable Luau code. The skill documentation explicitly warns that backdoors are common in free models and advises manual inspection.
[DYNAMIC_EXECUTION]: The agent uses execute_luau to generate and run code dynamically within the Roblox Studio process to inspect state and modify the DataModel.
[DATA_EXFILTRATION]: The screen_capture tool allows the agent to take screenshots of the Roblox Studio viewport, which may expose sensitive code, UI designs, or project metadata.

roblox-studio-mcp