Agentic AI tools (e.g. Claude Code, OpenAI Codex, OpenCode) are LLM-based assistants that can run commands, edit files and submit jobs on your behalf. They may be used on our HPC systems under the following conditions.
- You are responsible. Agents run under your personal account and act on your behalf. Every command they run is your command, and you are bound by the Sigma2 Terms of Use and the Acceptable Use Policy regardless of whether you or the agent typed it. This applies equally to agents running on your own machine that execute commands on our systems over SSH.
- On login nodes: one supervised session. Agents follow the login node rules like any other process. One interactive agent session per user is allowed, provided you are attached to it and actively watching it, and the agent itself only does light work. Detaching the session and leaving (tmux, screen, nohup, background jobs) is unsupervised use and is not allowed on login nodes. Anything the agent starts (compilation, tests, data processing) counts as the agent’s work.
- Be gentle with shared services. Instruct your agent accordingly: no tight polling of the queue system (e.g. squeue/sacct), no recursive find/grep over large directories on the parallel filesystem and no bursts of test jobs. This also applies to the jobs the agent submits.
- You are responsible for what you share with the provider. Agents send code, files, command output and error messages to an external provider. Before you let an agent read something, make sure it may be sent to the AI-service provider. Share personal, sensitive, or otherwise protected data only if you have a legal basis and your institution has an agreement with the provider that covers it. For other project data, make sure the project leader and any agreements covering the data allow it.
- Do not expose other users to the agent. On a shared system you can see usernames, jobs and processes that are not yours to pass on. Restrict the agent to your own data and jobs (for example squeue --me, sacct without -a and ps -u $USER), and do not let it list other users’ processes, jobs or directories without restriction (for example who, top, ps aux, squeue or ls in other users’ directories).
- Keep your credentials to yourself. Never give passwords, SSH keys or tokens to an agent running on a service you do not control (web chatbots with tool use, cloud IDEs). Do not open the system to such a service via MCP servers, reverse tunnels or similar; this is equivalent to sharing your credentials. Do not let the agent read ~/.ssh or files containing secrets. Running an agent on your own laptop or workstation, where your SSH keys already live, is fine and often preferable.
- Your tools, your support. NRIS does not install or support agents. For tool issues, contact the tool’s vendor.
- Enforcement. Processes or jobs that disturb other users or services may be stopped without warning. Breaches of this policy are handled under the Acceptable Use Policy. Depending on severity, this can include restricting or suspending your access and informing your project leader and, where relevant, your institution.
This document may change. You are responsible for checking that your use is in line with the current version. Questions and feedback: contact@sigma2.no
Last updated: 2026-09-29