The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions to install dependencies from the official Facebook Research GitHub repository and download pre-trained model weights from HuggingFace. These targets are well-known and trusted sources.
[COMMAND_EXECUTION]: The troubleshooting and advanced usage guides contain instructions for executing system commands, including package installation via apt-get and experiment management using dora run. These are standard administrative actions for the skill's domain.
[INDIRECT_PROMPT_INJECTION]: The skill describes several deployment scenarios that process untrusted text inputs, creating a vulnerability surface for indirect prompt injection.
Ingestion points: User-provided text prompts are ingested through a Gradio interface (SKILL.md), a FastAPI request body (references/advanced-usage.md), and a LangChain tool input (references/advanced-usage.md).
Boundary markers: The code examples do not implement delimiters or specific instructions to isolate the user-provided prompts from the agent's internal logic.
Capability inventory: The skill scripts perform file system writes using torchaudio.save and scipy.io.wavfile.write, and initiate network connections to fetch model parameters.
Sanitization: There is no evidence of input validation, filtering, or escaping for the music and sound descriptions before they are passed to the generation models.

audiocraft-audio-generation