review-music

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

• Third-party content exposure detected (medium risk: 0.30). In src/analysis/llm_analyzer.py, runtime transcription/feature text derived from outsider-provided audio input (local file path or --youtube URL downloaded audio; lyrics text included) is formatted into the LLM prompt (features['lyrics']['text'] → format_features_for_prompt → create_analysis_prompt) and then sent to the LLM via the /assistant gateway/scillm completion, so attacker-controlled free text can reach the model through the lyrics field.