The Agent Skills Directory

[PROMPT_INJECTION] (LOW): Vulnerable to indirect prompt injection. The skill transcribes audio from external, untrusted Douyin videos and uses the result to generate summaries, which could allow malicious instructions in the video to influence the agent.
Ingestion points: output.txt (transcription of user-provided Douyin video content).
Boundary markers: Absent; the summary template interpolates transcription text without specific delimiters or instructions to ignore embedded commands.
Capability inventory: Subprocess execution (ffmpeg, whisper-cli), file system writes, and network access (Feishu API).
Sanitization: None; transcribed text is processed directly by the LLM.
[DATA_EXFILTRATION] (LOW): The skill makes network requests to douyin.com and open.feishu.cn, which are not on the trusted whitelist for network operations.
Evidence: scripts/download_audio.sh (Douyin) and references/feishu-sync.md (Feishu API tokens and data submission).
[EXTERNAL_DOWNLOADS] (LOW): Downloads binary data from user-provided URLs and a model from Hugging Face.
Evidence: SKILL.md documents downloading a Whisper model from Hugging Face (a trusted organization).
[COMMAND_EXECUTION] (LOW): Orchestrates local binaries (ffmpeg, whisper-cli) using shell scripts to process media files.
Note on Scanner Alerts: The scanner alert for 'url.in' is a false positive likely caused by the string url.includes in the documentation. The 'Remote Code Execution' alert for douyin.com is also a false positive, as it refers to processing downloaded audio with ffmpeg rather than executing untrusted code.

douyin-video-summary