Responsible Use of LLMs
LLM tools are useful, and they are also able to read your files, run commands with your permissions, and state wrong answers confidently. These habits apply whether you use the ORC gateway, your own llama.cpp server, or any other service.
Mason's Responsible Use of Computing policy applies in full.
What goes into a prompt
No keys or passwords, ever. This includes your ORC virtual key, SSH private keys, .env files, and database credentials. Once a secret is sent, assume it is stored somewhere. If you send one by mistake, rotate it: for an ORC key, email orchelp@gmu.edu.
No restricted data. Don't send student records (FERPA), health information (HIPAA), export-controlled material, or data covered by an IRB protocol or a data use agreement (DUA) unless those agreements explicitly allow it. When in doubt, ask your PI or your IRB before you prompt.
Check what comes back. Verify code, numbers, and citations. A model sounds equally sure when it is wrong, and it can invent references that don't exist.
Running an agent
Agent tools such as Cline, Claude Code, and VS Code's agent mode can edit files and run commands on your behalf.
Never turn on auto-approve. Make the agent ask before every file write, shell command, and package install, and approve one step at a time.
Give it the project folder, not everything. An agent reads and writes with your permissions. Open only the project you're working on; don't point it at your home directory, /scratch, or shared project storage.
Read the diff before you accept it. Agents often change more than you asked for. Look at what changed, then run it.
Keep heavy work off the login nodes. If an agent runs builds, tests, or training on Hopper, it must do so on a compute node inside a Slurm allocation, like any other work.
Credit and academic integrity
Follow your course, department, and publication venue's rules on AI assistance, and disclose it where they require. You are responsible for any code or text you submit, whoever or whatever wrote the first draft.