The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill enables the agent to process untrusted external content, creating an attack surface for indirect prompt injection.\n
Ingestion points: Untrusted data is ingested via vision describe (images/URLs), speech transcribe (audio files), and search query (web results) in SKILL.md.\n
Boundary markers: No specific delimiters or instructions are provided to the agent to disregard instructions within these external inputs.\n
Capability inventory: The skill possesses capabilities for reading local files, writing files, and communicating with external APIs.\n
Sanitization: No sanitization of ingested content is defined in the instructions.\n- [DATA_EXFILTRATION]: The skill provides the capability to read local files and transmit their contents to the MiniMax API for processing, which constitutes a potential data exfiltration vector if the agent is misdirected.\n- [DYNAMIC_EXECUTION]: In h3-video/SKILL.md, the skill directs the agent to build local project artifacts using bun run build and execute them via node ./dist/mmx.mjs, which involves the execution of runtime-compiled code.\n- [CREDENTIALS_UNSAFE]: The skill manages MiniMax API keys and persists them in the ~/.mmx/config.json configuration file as part of its authentication workflow.\n- [EXTERNAL_DOWNLOADS]: The skill documentation references the installation of the mmx-cli package from the npm registry to provide its core functionality.

mmx-cli