The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the Hugging Face Hub, which is a platform where users can upload unverified models, datasets, and documentation. This content could be crafted to include instructions that manipulate the agent's behavior.
Ingestion points: Tools such as model_search, dataset_search, space_search, paper_search, hub_repo_details, and hf_doc_fetch retrieve content (e.g., README files, paper abstracts, and documentation) from user-controlled repositories on the Hugging Face Hub.
Boundary markers: The provided tool usage examples do not include explicit boundary markers (such as XML tags or specific delimiters) to isolate external content from the agent's instructions. However, the 'Anti-Patterns' section provides a warning to the agent to not treat tool responses as trusted instructions.
Capability inventory: The skill has access to high-privilege tools such as hf_jobs (which executes Python scripts or shell commands) and dynamic_space (which invokes external Gradio applications).
Sanitization: There is no documentation of automated sanitization or validation of the retrieved content before it enters the agent's context.
[REMOTE_CODE_EXECUTION]: The hf_jobs tool provides a mechanism for running arbitrary Python code and shell commands on remote Hugging Face infrastructure. While this is a standard feature for this vendor's compute services, it constitutes a remote code execution capability that must be monitored.
[COMMAND_EXECUTION]: The skill demonstrates the execution of shell commands, such as pip install and python train.py, within a remote environment via the hf_jobs tool, enabling arbitrary command execution in the target compute instance.
[TIME_DELAYED_CONDITIONAL]: The hf_jobs tool supports a scheduled uv operation that uses cron syntax to schedule the execution of scripts. This capability can be used to execute code at a later time, which is a pattern commonly used for persistence or delayed-execution workflows.

hf-mcp