read-xiaoyuzhou-article

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

• Third-party content exposure detected (high risk: 0.75). Yes—scripts/transcribe.sh fetches outsider-authored public web content from the user-supplied 小宇宙 episode URL (curl -s ... "$URL"), extracts the audio transcript text from that page’s media and then sends the resulting transcript (free-form text derived from that outsider content) into Groq/LLM calls via -F prompt="$PROMPT" and -F file="@$1" (and later into the Llama “polish” prompt via PROMPT_TMPL.format(text)), creating an indirect prompt-injection path.