The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (MEDIUM): The skill utilizes a build(code=...) tool to execute dynamically generated Python scripts at runtime. This 'Script generation + execution' pattern is inherently risky as it relies on the robustness of the sandbox environment to prevent unauthorized actions or resource exhaustion during the building process.
[PROMPT_INJECTION] (MEDIUM): The skill presents an Indirect Prompt Injection surface by transforming natural language parameters into executable Python logic. Evidence Chain: (1) Ingestion point: User-provided dimensions and geometry descriptions. (2) Boundary markers: No delimiters or explicit 'ignore instruction' warnings are present in the code generation examples. (3) Capability inventory: Python code execution via the build tool and Minecraft command execution. (4) Sanitization: The skill relies on stated sandbox limits (blacklisting 'import', 'def', etc.) which may be bypassable through sophisticated injection techniques.
[COMMAND_EXECUTION] (LOW): The skill generates and issues Minecraft console commands (e.g., '/setblock', '/sphere'). While this is functional, it constitutes an automated path for the agent to modify the environment state based on potentially untrusted input.

creating-shapes