The Agent Skills Directory

[Command Execution] (HIGH): The twitter_video_downloader.py script is vulnerable to command injection via the --urls parameter. The script splits the input string by commas and appends the resulting list directly to the yt-dlp command arguments. An attacker can provide flags such as --exec followed by a command (e.g., --urls "--exec,id") to execute arbitrary code on the host system.
[Credentials Unsafe] (MEDIUM): The downloader script requires a cookies.txt file containing sensitive browser session data for authentication. If this file is accessed by an attacker or if the environment is compromised, it could lead to full account hijacking.
[Data Exposure & Exfiltration] (MEDIUM): The transcribe_siliconflow.py script sends local audio and video files to an external third-party API (api.siliconflow.cn). This exposes potentially private user content to an external service provider.
[Indirect Prompt Injection] (MEDIUM): Transcription results from untrusted audio/video files are saved to Markdown without any sanitization or boundary markers. 1. Ingestion: Untrusted media files. 2. Boundary markers: Absent. 3. Capability: Writing to local files and potential downstream processing. 4. Sanitization: Absent. A maliciously crafted audio file could contain verbal instructions that, once transcribed, might influence or hijack the reasoning of an agent reading the output file.

advanced-video-downloader