podcast-personalization-research

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

• Third-party content exposure detected (high risk: 0.85). High: the skill uses WebSearch to find podcast/episode pages and then WebFetches the top results, ingesting outsider-authored show notes/transcripts/episode text (public web content) into the agent’s LLM context for quote/topic extraction.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly calls WebFetch at runtime to retrieve podcast episode pages (e.g., https://podcasts.apple.com/..., https://open.spotify.com/episode/..., YouTube/show-notes URLs) and injects extracted quotes and content into the agent's synthesized outputs, so fetched external page content directly controls the agent's prompts and outputs.