30. Security and Compliance#

Working on the Kempner AI cluster is a shared responsibility. The cluster is part of Harvard’s FASRC environment, so your account, your data, and the jobs you run are all subject to university and FASRC policy. This page summarizes the practices and rules that keep the cluster secure and your research compliant. It points to the authoritative Harvard and FASRC pages for the details, which take precedence over anything summarized here.

See also

The overarching policies live in the FASRC Acceptable Use Policy and Harvard Research Data Management and Policy Compliance. For cluster-specific expectations, see Cluster Usage Policies.

30.1. Prohibited activities#

The following activities are not allowed on the cluster.

  • Sharing credentials or accounts. Do not share your cluster credentials or let another person use your account. Cluster access is tied to an individual user and must not be shared; see Account and access security.

  • Bypassing SLURM. Do not bypass, or try to bypass, SLURM scheduling and resource-allocation policies. Access and allocate compute resources only through approved mechanisms.

  • Unauthorized remote access. Do not set up unauthorized remote-access mechanisms that bypass normal cluster authentication, obscure user identity, or circumvent SLURM resource allocation on compute nodes. This includes tunnels, proxies, remote-access services, and similar mechanisms set up to bypass cluster controls.

  • Scanning or probing. Do not scan or probe cluster systems, users, processes, directories, services, or networks without explicit authorization. This includes vulnerability scanning, port scanning, process enumeration, unauthorized access attempts, and exploring resources that you are not authorized to access.

  • Uncontrolled agentic workflows. Do not run unsupervised or insufficiently controlled agentic workflows at scale if they could interact with cluster infrastructure, networks, user data, credentials, or external systems in ways that create security risk. You remain responsible for the actions of automated and agentic workloads that run under your account; see Responsible use of AI tools and Before an unattended run.

  • Exposing the cluster to external networks. Do not expose cluster services or compute nodes directly to external networks without authorization. This includes opening ports, running public-facing servers, and creating network tunnels that circumvent approved access controls.

  • Storing or exposing secrets. Do not store or expose sensitive credentials in source code, shared directories, notebooks, logs, job scripts, or other locations that unauthorized users can access. Sensitive credentials include passwords, API keys, access tokens, SSH private keys, and other secrets.

  • Evading security controls. Do not disable, interfere with, or try to evade security monitoring, logging, access controls, or other security mechanisms.

  • Unauthorized or malicious activity. Do not use cluster resources for unauthorized or malicious activity. This includes attempts to compromise other systems, disrupt services, or exfiltrate data, and using resources for purposes that Harvard, Kempner Institute, or FASRC policies prohibit.

These rules do not restrict normal cluster work, such as running jobs with sbatch, salloc, or srun, connecting with SSH or VS Code Remote-SSH to a node where your job runs, serving a model on a compute node for your lab, or checking the queue with squeue.

If you are not sure whether an activity is allowed, or you need authorization for one, ask FASRC at rchelp@rc.fas.harvard.edu before you start.

30.2. Account and access security#

Your cluster account is tied to you personally, and you are accountable for everything done under it.

Warning

Never share your account or credentials with anyone, including labmates. Account sharing violates university policy: accounts found sharing credentials may be disabled or banned, access can be revoked, and further university action may follow. Authentication and access logs are actively monitored for suspicious activity and unusual login patterns, and user accounts have already been disabled in response to credential sharing. If someone needs access, they must request their own account. See Can I share my account?.

To keep your credentials safe:

  • Use a password manager. Harvard provides 1Password free to the community, and its use is encouraged.

  • Keep your two-factor (OpenAuth) device secure, and never share one-time codes. See the New User Checklist for setup.

  • Report a suspected account compromise immediately; see Reporting a concern.

Warning

Remote control of AI agents. Some AI agents let you drive a running session from a phone or a browser, including Claude Code’s Remote Control, codex remote-control, and the ChatGPT app’s remote connections. On the cluster, a session with remote control on can be steered from any device signed in to your AI account, without the FASRC login and two-factor authentication that normally protect your cluster account. Remote control bypasses normal cluster authentication, so do not turn it on for sessions on the cluster; see Prohibited activities and Remote control.

30.3. Data classification and what the cluster can host#

Harvard classifies data into five security levels. The level determines where the data may be stored and processed.

Level

Meaning

Examples

L1

Public

Published papers, open datasets

L2

Low-risk confidential

Most unpublished research data, non-sensitive code

L3

Confidential and sensitive

Many types of personal data, some regulated data

L4

High risk

Sensitive personal or regulated data

L5

Extremely sensitive

The most tightly controlled data

See the Harvard Data Classification Table for the authoritative definitions.

Warning

The Kempner AI cluster is rated for data up to Level 2 (L2) only. Do not place L3 or higher data on the cluster. Level 3 data must be handled in Harvard’s dedicated FASSE secure environment, which is rated for L3 only; Level 4 and above are not permitted on any FASRC system. For regulated (Level 4) data, use Harvard University Research Computing’s Regulated Data Services.

Warning

This means no PII, HIPAA/PHI, or other regulated high-risk data on the cluster, since Harvard classifies these Level 3 or higher (HIPAA/PHI and PII are Level 4). If you are unsure how your data is classified, confirm with your PI and Harvard security before uploading it.

30.4. Data ownership and handling#

Research data generated at Harvard is owned by the University; you are its steward, not its private owner. How you move and share it is governed by grant terms, data use agreements (DUAs), and your PI’s direction.

Warning

Do not move data out of a PI’s lab space, whether to another location on the cluster or to any external destination, without PI approval and confirmation that it complies with the relevant grant terms and data use agreements. Unauthorized copying or sharing of lab data can breach funding and legal obligations.

Note

Any required data use agreement must be in place before the data is stored on the cluster, and the PI is responsible for ensuring approved access stays compliant. For planning your data lifecycle, see Data Management Plan and Storage Options.

Note

Retention and deletion. Research records must be kept for the retention period Harvard and your sponsor require, generally at least seven years, and you may not independently delete or remove research data. Scratch storage is purged after 90 days, so it is not a retention solution; copy anything you must keep to persistent or archival storage. See the FASRC Research Data Retention and Deletion Policy and the Data Management Plan.

30.5. Responsible use of AI tools#

AI coding assistants and other generative AI tools are useful on the cluster, but their use must follow university guidance.

Note

Follow Harvard’s Generative AI Guidelines and the Provost’s guidance on using ChatGPT and other generative AI tools. In particular, do not paste confidential (L2 and above) data into public AI services.

Warning

Take extra care when running AI tools in automated or agentic mode, where the tool executes commands on its own. Before enabling autonomous execution:

  • Review the actions the agent proposes rather than approving them blindly.

  • Scope the agent to your own directories, and avoid destructive operations (for example, recursive deletes).

  • Ensure it cannot affect shared paths, other labs’ data, or cluster-wide resources.

An unsupervised agent can delete data or disrupt shared resources far faster than a person.

Note

Agentic tools are also exposed to prompt injection: untrusted content the agent reads, such as web pages, files, or datasets, can carry hidden instructions. Treat what an agent reads as data rather than commands, keep a human approving consequential actions, and give the agent least privilege. See OWASP’s Top 10 for Agentic Applications and, for running these tools here, Agent Security and Scoping.

Note

Scope an autonomous agent to its task. An agent runs with your account’s authority, which usually exceeds what any single task needs, so reduce its effective authority to match the task’s purpose using controls that do not depend on the model. When scope is ambiguous, prefer stopping for human review over letting the agent widen its own scope. See Agent Security and Scoping for how to do this.

Warning

Prohibited AI under the NDAA. Harvard’s NDAA 2026 Guidance and Prohibition on use of certain Artificial Intelligence prohibits anyone performing work under a U.S. Department of Defense (DoD) contract from using AI developed by DeepSeek, its parent company High Flyer, or affiliated entities, in any activity related to that contract. The prohibition covers faculty, staff, students, and visitors working on the contract, and it applies to self-hosting the open-weight DeepSeek models on the cluster as well as the hosted service. If your work is, or may be, under a DoD contract, do not use these tools, and contact the Office of the Vice Provost for Research if you are unsure whether the policy applies to you.

30.6. External data and network conduct#

Datasets you download or scrape from external providers come with terms of use, and the whole cluster shares a small pool of public IP addresses.

Warning

Respect the terms of service and rate limits of any external data source. Aggressive downloading or scraping that violates a provider’s terms can get the cluster’s shared public IP blocked, breaking access for every user, and may breach the provider’s agreement. When in doubt, throttle your requests and read the source’s terms first.

See the FASRC Acceptable Use Policy for the full expectations on network conduct.

30.7. Software licensing#

You may install software relevant to your research, provided you comply with the FASRC Acceptable Use Policy and the licensing terms of each package.

Note

FASRC does not purchase software for labs. Commercial or licensed software must be acquired by your lab, department, or school (for example, through HUIT software licensing), and must have a Linux-compatible license to run on the cluster. See Installing Software and FAS licensed software.

Respect the license of every dataset, model, and library you use, including non-commercial and attribution clauses. This is both a legal and a compliance requirement.

30.8. Compliance and grant obligations#

Your work is also bound by the requirements of whoever funds and governs it.

  • Funding agencies. Follow the data management, sharing, and access requirements of your grants (for example, NIH and NSF data policies). These often flow into your Data Management Plan.

  • Retention periods. Keep research data for the retention period your sponsor and Harvard require. Funding agencies set their own minimums, and Harvard’s longer retention period supersedes shorter federal minimums. See the Data Management Plan.

  • Data use agreements. Honor every DUA and contract that applies to data you use, as directed by your PI.

  • University policy. You are responsible for knowing and following applicable Harvard Information Security and Research Data Security policies.

Reporting a concern

If you suspect a security incident, a compromised account, or a policy violation, report it promptly. Contact FASRC at rchelp@rc.fas.harvard.edu and, for security incidents, Harvard’s Information Security office at ithelp@harvard.edu. Acting quickly limits the impact on you, your lab, and the wider cluster.

Note

If an autonomous agent behaves unexpectedly, the reporting guidance above applies, with these agent-specific steps:

  1. Stop the agent or job first. Cancel the SLURM job or terminate the process before investigating.

  2. Preserve run metadata. Save the SLURM job ID, logs, and relevant output, without copying secrets into new files.

  3. Rotate exposed credentials. If the agent may have read or transmitted API keys, tokens, or other secrets, revoke and rotate them.

  4. Escalate when appropriate. If shared data, other users’ resources, cluster stability, or a data-use agreement may be affected, report it to your PI and to FASRC.

  5. Do not use the same agent to investigate. An agent that may have been misdirected should not assess its own behavior.