llm-prompting-guide

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The required runtime workflow ingests two user-provided free-text fields (“Caption” and “Lyrics”, plus optional “negative_caption”) directly into the MiniMax Music 3 text encoder (T2T→caption/lyrics tokenization) where the model reads that arbitrary text.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill includes shell commands that clone https://github.com/MiniMax-AI/MiniMax-Music3.git at runtime (git pull || git clone ...), which would fetch prompt/template files that directly control the model prompts and are suggested as a dependency for the skill.