The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen in several examples to launch external tools such as torchrun, accelerate, and the vLLM server. These are documented with clear security warnings advising the use of hardcoded argument lists to prevent command injection.
[DYNAMIC_EXECUTION]: The skill facilitates the execution of dynamically generated or untrusted code using modal.Sandbox, which is designed for isolated execution. It also supports running Python functions during the container image build process via Image.run_function and supports dynamic loading of local modules via add_local_python_source.
[INDIRECT_PROMPT_INJECTION]: The skill presents vulnerability surfaces where the agent might process untrusted data.
Ingestion points: Data is fetched from arbitrary URLs in the web scraping example (references/examples.md) and accepted via JSON request bodies in web endpoint examples (references/web-endpoints.md).
Boundary markers: The provided examples do not explicitly demonstrate the use of delimiters or prompt instructions to ignore embedded commands in the fetched content.
Capability inventory: The skill provides access to powerful capabilities including subprocess.run for shell execution, modal.Sandbox for running arbitrary code, modal.Volume for persistent file system access, and modal.Secret for managing credentials.
Sanitization: The skill provides security guidance, recommending the use of modal.Sandbox for untrusted code execution and using fixed argument lists for subprocess calls. It also includes an example of validating input identifiers (e.g., doc_id) using character allowlists.
[EXTERNAL_DOWNLOADS]: The skill downloads the modal SDK and various machine learning libraries (e.g., torch, transformers, vllm) from standard package registries. It also fetches pre-trained models from Hugging Face Hub, which is a well-known service. All downloads target trusted organizations or well-known services.

modal