create-song

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

• Third-party content exposure detected (high risk: 1.00). The SKILL.md Step 2 "Research the Style" explicitly instructs the agent to run WebSearch and then WebFetch on open/public sites (e.g., WhoSampled, Hooktheory, Splice, Reddit production threads) and to extract and use full page content to determine chord progressions, production details, and rendering decisions, so it clearly ingests untrusted third‑party content that can materially influence tool use and next actions.

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

• Potentially malicious external URL detected (high risk: 0.70). The skill explicitly instructs using WebSearch/WebFetch at runtime to retrieve full pages from external sites (e.g., https://hooktheory.com and https://whosampled.com) and to extract page content that is then injected into the agent's context to drive code generation and rendering decisions, so these external URLs directly influence prompts/runtime behavior.