hugging-face-datasets

Warn

Audited by Gen Agent Trust Hub on Aug 9, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands using the uv package manager to run local scripts (e.g., uv run scripts/dataset_manager.py).
  • [EXTERNAL_DOWNLOADS]: The scripts use PEP 723 inline dependency management, which causes the uv tool to automatically download and install Python packages from external registries when the scripts are executed.
  • [DATA_EXFILTRATION]: The skill is designed to transmit local data and metadata to the Hugging Face Hub using the push-to and add_rows commands, utilizing a user-provided HF_TOKEN for authentication.
  • [COMMAND_EXECUTION]: Recommended usage patterns involve shell command substitution (e.g., $(cat system_prompt.txt)) to pass local file contents as command-line arguments, which provides a path for reading local files into the agent's context.
  • [COMMAND_EXECUTION]: The sql_manager.py tool provides an interface for executing arbitrary SQL queries via DuckDB, which includes capabilities to read from local files and remote URLs.
  • [PROMPT_INJECTION]: The skill processes untrusted data from the Hugging Face Hub via the hf:// protocol, which constitutes a surface for indirect prompt injection if the datasets contain malicious instructions targeting the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 9, 2026, 09:28 PM
Security Audit — agent-trust-hub — hugging-face-datasets