The Agent Skills Directory

[COMMAND_EXECUTION]: The runMmx function in minimax-mcp-server.js executes CLI commands via child_process.execSync using shell string interpolation. The implementation only escapes double quotes and fails to escape shell-active characters like $, backticks, and backslashes. This allows for arbitrary command substitution and code execution if a user-provided argument (such as a prompt, message, or file path) contains malicious shell syntax.
[EXTERNAL_DOWNLOADS]: The skill requires the mmx-cli tool to be installed globally on the system. This is an external dependency that is not part of the standard environment or a verified vendor repository.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface by ingesting and processing untrusted data from external sources without proper isolation.
Ingestion points: Untrusted data enters the agent context through minimax_search (web content), minimax_vision_describe (image content), and minimax_text_chat (external model responses).
Boundary markers: The skill does not employ delimiters or system instructions to separate external content from the agent core logic in SKILL.md or the server script.
Capability inventory: The MCP server can execute arbitrary shell commands and perform file-system operations (write/download) via its underlying CLI wrapper.
Sanitization: No input validation, escaping, or filtering is performed on the data retrieved from external tools before it is passed to the agent or used in subsequent tool calls.

minimax-rules