The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/transcribe.py to invoke yt-dlp for downloading media files from user-provided URLs. It also executes a local summary.py script for post-transcription processing. These calls use argument lists rather than raw shell strings, which mitigates common shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: As part of its core functionality, the skill downloads media content from third-party platforms (such as YouTube, Bilibili, and Xiaoyuzhou) using the yt-dlp utility. The downloaded files are stored in a local temporary directory before being uploaded to Alibaba Cloud.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of audio and video files which are then converted to text. This transcription process represents a potential attack surface where instructions embedded in the media content could attempt to influence the agent's behavior during subsequent analysis or summarization tasks.
Ingestion points: Media files provided via local paths or remote URLs in scripts/transcribe.py and scripts/tingwu.py.
Boundary markers: The output is structured into Markdown, but there are no specific delimiters or warnings to the agent to ignore instructions embedded within the transcribed text.
Capability inventory: The skill has capabilities to write files to the local system, access Alibaba Cloud APIs, and execute internal helper scripts.
Sanitization: There is no evidence of filtering or sanitizing the transcribed text for malicious command patterns before it is provided to the agent.
[CREDENTIALS_SAFE]: The skill requires Alibaba Cloud credentials and session cookies. It follows security best practices by instructing users to store sensitive information in a local .env file and providing .gitignore entries for both the environment file and the session cookie storage, ensuring they are not committed to version control.

tingwu-asr