magic-data-lifecycle
Warn
Audited by Gen Agent Trust Hub on Jul 3, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The pipeline_template.py file uses importlib to dynamically load and execute code from a configuration file path provided via a command-line argument, enabling the execution of arbitrary Python logic from external files.
- [REMOTE_CODE_EXECUTION]: Skill instructions guide the agent to generate and save executable Python code for pipelines based on conversation history and user requirements.
- [EXTERNAL_DOWNLOADS]: The skill provides patterns for downloading datasets from HuggingFace using the official huggingface_hub library.
- [DATA_EXFILTRATION]: Examples within the skill demonstrate uploading processed data to remote repositories on HuggingFace Hub, which serves as a legitimate channel for data export.
- [PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data (CSV, JSONL, Parquet) through a multi-stage pipeline that includes LLM-based synthesis, creating an attack surface where malicious data content could influence the agent's behavior. Evidence found in SKILL.md (ingestion) and pipeline examples (capability).
Audit Metadata