The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The install.sh script downloads and executes a script from pixi.sh via a pipe to bash. While pixi is a well-known tool, this installation pattern represents a potential remote code execution vector.\n- [CREDENTIALS_UNSAFE]: The bilibili_subtitle/preflight.py script reads the ~/BBDown.data file to check for Bilibili session cookies (SESSDATA). This file contains sensitive authentication data used to access the Bilibili account.\n- [COMMAND_EXECUTION]: Several modules, including bbdown_client.py, chunker.py, and transcribe_agent.py, use the subprocess module to execute external binaries like BBDown, ffmpeg, and ffprobe using arguments derived from user input.\n- [EXTERNAL_DOWNLOADS]: The install.sh script uses the GitHub CLI to download nightly versions of the BBDown tool from the nilaoda/BBDown repository.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from Bilibili video subtitles and transcripts. This data is ingested in subtitle_loader.py and passed to LLM agents for proofreading and summarization in agents/proofread_agent.py and agents/summarize_agent.py, which lack strong boundary delimiters between instructions and the untrusted external content. The skill has capabilities to write files and execute system commands.

bilibili-subtitle