The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/setup-woosh.sh installs the uv tool by downloading a shell script from https://astral.sh/uv/install.sh and piping it directly to sh. This pattern is highly risky as it executes remote code without prior verification.
[REMOTE_CODE_EXECUTION]: The setup scripts (setup-mmaudio.sh/ps1 and setup-woosh.sh/ps1) clone third-party repositories from GitHub (hkchengrex/MMAudio and SonyResearch/Woosh) and execute installation commands like pip install -e . and uv sync. These repositories are outside of the trusted vendor scope and their contents are executed automatically.
[COMMAND_EXECUTION]: Instructions in AGENTS.md direct the AI agent to automatically execute the setup and installation scripts (setup-woosh and setup-mmaudio) when local processing is required, bypassing the need for explicit user confirmation before system-level changes occur.
[EXTERNAL_DOWNLOADS]: The skill initiates the download of approximately 5GB of external model weights from GitHub and HuggingFace during the automatic setup process.
[DATA_EXFILTRATION]: The local server scripts/mmaudio-server.py includes a /generate endpoint that accepts an arbitrary video_url. The server uses urllib.request.urlretrieve to download the content of this URL to the local filesystem. This creates a Server-Side Request Forgery (SSRF) vulnerability that could be used to probe internal network services accessible to the agent's environment.

sonic