T09 · Insecure Skill Coding Practices
Error
- Location
- scripts/asr.py:358
- Finding
- Default ASR selection can upload video audio to a cloud service without per-invocation consent<![CDATA[ ## Vulnerability Details **File Location**: `scripts/download_video.py:32-39`, `scripts/download_video.py:100-108`, `scripts/asr.py:94-101`, `scripts/asr.py:358-378`, and `scripts/asr.py:492-496` **Vulnerability Type**: Implicit transmission of potentially sensitive media to a third-party service **Risk Level**: High ### Complete Code Snippet ```python # scripts/download_video.py:32-39 parser.add_argument( "--asr", choices=("auto", "whisper_cpp", "whisper", "siliconflow", "none"), default="auto", help=( "Audio transcription backend. Default: auto " "(whisper.cpp > openai-whisper > SiliconFlow)" ), ) ``` ```python # scripts/download_video.py:100-108 if not args.metadata_only and args.asr != "none" and result.get("video_path"): asr_result = run_asr( Path(result["video_path"]), Path(result["output_dir"]), backend=args.asr, model=args.asr_model, language=args.asr_language, prompt=args.asr_prompt, max_seconds=args.asr_max_seconds, ) ``` ```python # scripts/asr.py:94-101 if selected_backend == "siliconflow": transcribe_with_siliconflow( audio_path, transcript_path, siliconflow_json_path, model=resolved_model, ) raw_json_path = siliconflow_json_path ``` ```python # scripts/asr.py:358-378 api_key = os.environ.get("SILICONFLOW_API_KEY") if not api_key: raise RuntimeError("SILICONFLOW_API_KEY is required for ASR backend 'siliconflow'.") body, content_type = _multipart_body( fields={"model": model}, file_field="file", file_path=audio_path, file_content_type="audio/mpeg", ) request = Request( SILICONFLOW_TRANSCRIPTION_URL, data=body, headers={ "Authorization": f"Bearer {api_key}", "Content-Type": content_type, }, method="POST", ) try: with urlopen(request, timeout=600) as response: response_text = response.read().decode("utf-8", errors="replace") ``` ...[truncated 2465 chars]
- Remediation
- <![CDATA[ ## Remediation Suggestions 1. Make `auto` strictly local-only: - Check whisper.cpp and OpenAI Whisper. - If neither is available, return a clear error or pending result. - Do not select a network backend automatically. 2. Require explicit cloud authorization, such as: ```bash --asr siliconflow ``` or: ```bash --allow-cloud-asr ``` 3. Before uploading, display or return a structured disclosure stating: - The destination hostname. - The exact file being uploaded. - Whether the full audio or only a limited segment will be sent. - That a third party will process the content. 4. In Agent environments, require a per-task consent signal rather than inferring consent from the presence of an environment variable. 5. Consider adding a global offline mode that blocks every non-platform network request, including cloud ASR. 6. Retain the existing API-key redaction and extend it to unexpected exceptions and diagnostic output. ]]>