The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill facilitates the deployment and execution of Python code on Modal's serverless infrastructure via the .remote() and .deploy() methods. This is the intended primary purpose of the platform.
[COMMAND_EXECUTION]: Documentation and examples include the use of subprocess.run and subprocess.Popen to launch distributed training scripts (e.g., Hugging Face Accelerate) and web servers (e.g., vLLM) within cloud containers. The skill includes specific security notes warning developers to use hardcoded argument lists and avoid constructing commands from unsanitized user input.
[DYNAMIC_EXECUTION]: The skill describes the use of modal.Sandbox for programmatically executing dynamically generated or untrusted code in isolated containers with restricted network access. It also details build-time execution methods like Image.run_commands and Image.run_function used to configure container environments.
[INDIRECT_PROMPT_INJECTION]: The skill enables the creation of web endpoints that process external data, creating a potential surface for indirect injection if not handled correctly.
Ingestion points: FastAPI/ASGI web endpoints and sandbox filesystem writes (sb.filesystem.write_text) in SKILL.md and references/web-endpoints.md.
Boundary markers: The skill suggests using network egress allowlists and modal.Sandbox for isolation when handling untrusted workloads.
Capability inventory: Containers have access to GPUs, persistent storage (Volumes), and the ability to trigger further remote function calls.
Sanitization: The documentation advises validating inputs against allowlists and passing untrusted values as data or environment variables rather than command-line arguments.

modal