transcript-repurposer

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

• Third-party content exposure detected (high risk: 0.85). Outsider-authored free text (the source video’s transcript) is ingested into the agent’s LLM context via Phase 0/1 when the user runs the local CLI on an external URL (e.g., YouTube/Instagram/TikTok) and then triggers “Repurpose the latest from my inbox,” causing the skill to read the transcribed text from Documents\Claude\Skills\_inbox\transcript-*.txt and pass it through Phases 1–8.

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

• Potentially malicious external URL detected (high risk: 0.80). The local transcription CLI calls Deepgram at https://api.deepgram.com/v1/listen at runtime (scripts/transcribe_local.py), and the returned transcript is injected into the agent's source_text/context that the skill uses to generate prompts and outputs, so this is a runtime external dependency that can directly control the agent input.