AI Tool Guidelines on Monsoon

Data Classification

When using Monsoon, do not paste or upload sensitive data into public AI tools! This is the most important thing to keep in mind! Some examples of sensitive data include (but aren’t limited to) IRB-protected human subjects’ data, CUI, export-controlled data (ITAR/EAR), unpublished results, proprietary datasets, or anything with PII. Once data enters a public AI service, you no longer have control over it. If your research is federally funded, check your data management plan as some grant agreements restrict where data can be processed. If you are unsure how sensitive your data is, please refer to the NAU data classification page.

Least Privilege

Do not give an AI tool broad access to your data or script directories. Instead, only grant the access required for the specific task at hand. Unintended damage, accidental deletion, overwriting, or data exposure is possible when AI tools have more access than they need.

Agentic / IDE-Integrated Tools

If you’re using AI-enabled IDE extensions like Claude Code, GitHub Copilot, or Cursor via Remote-SSH into Monsoon, be aware that the extension can initiate outbound connections from within the cluster, potentially tunneling data to external AI services without you realizing it. Our recommendation is to use these tools to write your code locally on your laptop, then transfer your code to Monsoon manually.

LLM inference is very resource intensive and will impact other users if models are running on login nodes. If you want to run a locally-hosted AI model on Monsoon, that needs to go through a Slurm job like any other compute workload.

Protecting Your Credentials

Never give an AI tool access to your SSH keys, API tokens, .env files, or any credentials stored on Monsoon. Don’t paste cluster passwords or API keys into AI chat interfaces for debugging help, and scrub credentials from any code before sharing it with an AI tool.

Reviewing and Running AI-Generated Code

Always review scripts generated by AI tools before submitting them as Slurm jobs. Make sure you understand what every line does, especially anything touching file paths, data directories, or network calls. Be extra cautious with anything that uses rm, rsync, scp, or bulk data operations.

Sensitive Work

If you’re working with sensitive datasets and still want AI assistance, consider using a locally-run, offline model (e.g., Llama or Mistral running entirely within a Slurm job on Monsoon). These don’t make outbound connections and keep your data entirely on-premises. You can ask us to help you set that up if it’s a good fit for your workflow by contacting us. We plan to have a platform in the future that will make this easier for researchers.