universal-data-analyst-en

Performs automated, LLM-driven data analysis including loading, validation, method selection, script generation, execution, and comprehensive reporting for d...

This skill is coherent with its stated purpose, but it generates Python analysis scripts via LLM prompts and can execute those scripts locally. Before installing or running: 1) Do NOT run this on sensitive or production systems without reviewing generated scripts first. 2) Inspect any generated analysis_script.py for network, subprocess, or filesystem operations (look for imports like requests, socket, subprocess, os.system, eval/exec, urllib, ftplib, paramiko). 3) Prefer running the orchestration and script execution inside an isolated environment (ephemeral VM, container, or sandbox) with limited network and file access. 4) If you will call external LLMs, keep API keys separate and only use trusted endpoints; the skill does not manage credentials. 5) Consider using the human-in-the-loop mode (generate prompts and scripts but manually review/execute) rather than fully autonomous execution. If you want me to, I can: (a) scan the full repository for occurrences of subprocess/requests/os.system/eval/exec/network endpoints, or (b) point to specific lines/functions to review before running.

Type: OpenClaw Skill Name: universal-data-analyst-en Version: 1.0.3 The skill bundle implements a 'Universal Data Analyst' that uses LLMs to generate and execute Python code. The core risk lies in `orchestrator.py`, which uses `subprocess.run` to execute scripts generated by an LLM based on user-provided data and 'intent' strings, creating a significant Remote Code Execution (RCE) surface via prompt injection. Additionally, `layers/data_loader.py` supports loading data from SQL databases via arbitrary connection strings and executing SQL queries. While these capabilities are aligned with the stated purpose of the tool and no explicit evidence of malicious intent (such as hardcoded exfiltration URLs or credential theft) was found, the high-risk execution model and broad data access permissions warrant a suspicious classification.