The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a vulnerability surface for indirect prompt injection. It retrieves data from the Roblox Studio environment (including script source, object hierarchies, and properties) which could contain adversarial instructions placed by a third party. This data is then processed by the agent which has high-privilege capabilities.
Ingestion points: Tools such as script_read, inspect_instance, search_game_tree, and console_output (defined in references/tool-reference.md) ingest untrusted content from the active Roblox place.
Boundary markers: The skill does not employ prompt boundary markers or "ignore instructions" wrappers when processing the retrieved data.
Capability inventory: The skill provides high-privilege tools including execute_luau (arbitrary code execution), multi_edit (file writing), and insert_from_creator_store (asset ingestion), all found in references/tool-reference.md.
Sanitization: There is no evidence of sanitization or filtering applied to the external content retrieved from Studio before it is returned to the agent context.
[DYNAMIC_EXECUTION]: The execute_luau tool (described in references/tool-reference.md) allows the agent to run arbitrary Luau code strings within the Roblox Studio process. While this is a documented feature for developer productivity, it constitutes a dynamic execution vector.
[COMMAND_EXECUTION]: The installation instructions in references/setup-and-connection.md provide shell commands for starting the MCP server on Windows (cmd.exe /c %LOCALAPPDATA%\Roblox\mcp.bat) and macOS. These are necessary for the skill's operation but involve executing local system commands.

roblox-mcp