The Agent Skills Directory

Remote Code Execution and Command Execution: The skill provides a set of tools to execute code on Hugging Face compute infrastructure. While this is the primary purpose of the compute-focused features, it allows for the execution of arbitrary scripts and Docker containers.
Evidence: The hf_jobs tool supports operation="uv" for running Python scripts with specified dependencies (e.g., torch) and operation="run" for executing commands within specific Docker images (e.g., pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime).
Indirect Prompt Injection Surface: The skill ingests untrusted data from external sources, which could potentially contain instructions that influence the agent's behavior during a multi-step workflow.
Ingestion points: The hub_repo_details tool fetches repository READMEs, and hf_doc_fetch retrieves external documentation from the Hugging Face website. paper_search also retrieves content from scholarly papers.
Boundary markers: The skill instructions do not explicitly define delimiters for this external content, though the AI agent's underlying safety layers typically handle these boundaries.
Capability inventory: The agent has high-privilege capabilities including file system operations and network access through the hf_jobs and dynamic_space tools.
Sanitization: No explicit sanitization logic for ingested text is described in the skill definition.
Persistence Mechanisms: The skill includes functionality to schedule recurring compute tasks, which acts as a persistence mechanism within the Hugging Face infrastructure.
Evidence: The hf_jobs tool supports a scheduled uv operation using standard cron syntax for periodic execution.
Dynamic Execution: The skill facilitates the discovery and invocation of dynamic AI tools (Gradio Spaces) which may have their own execution environments.
Evidence: The dynamic_space tool allows for viewing parameters and invoking operations on external Gradio applications hosted on Hugging Face Spaces.

hf-mcp