Overview · RCC ClusterDocs

Off-site coding agents must not receive RCC research data

A coding agent—an AI tool that writes, tests, and revises code—can save a great deal of time. You can describe the table, plot, report, or workflow you want and ask it to build reproducible code.

But RCC projects may contain patient-derived, genetic, confidential, unpublished, or otherwise protected material. You must not paste, upload, or otherwise make real or pseudonymised RCC project data available to an off-site coding agent. Sending a prompt or file to such an agent is data processing and disclosure outside the controlled RCC project environment.

Health and genetic data are specially protected. Article 9 of the GDPR starts from a prohibition on processing these data unless a specific exception and safeguards apply. German research provisions such as section 27 BDSG impose necessity and protective measures; they are not blanket permission to disclose research data to another service. Medical confidentiality may impose additional duties under section 203 StGB.

An agent being licensed, commercially available, institutionally provided, or described as “approved” does not authorize it to receive RCC data. A user cannot create the required legal basis, processor arrangement, safeguards, or third-country transfer basis by opening an account or accepting a service's terms. Any exceptional off-site processing arrangement must be established and explicitly authorized by the responsible institution for that exact service and purpose. Do not infer such authorization from this documentation.

There are two practical routes.

Route A: use a coding agent inside RCC

Use an RCC-internal coding agent when it is available for your project. It works inside the RCC environment under your normal project access. The real data stays in RCC while the agent prepares code, requests bounded cluster computation, and writes the permitted results back to the project.

your request + real project data
              |
              v
     RCC-internal coding agent
              |
              v
       RCC workers -> project results

The agent does not become an administrator and does not gain access to another project. You still review important actions, and substantial computation runs through Slurm.

Read how the RCC-internal coding-agent route works and what it can do.

Route B: give an off-site coding agent synthetic data only

An off-site coding agent can still write useful code without seeing the real dataset. Give it a small coding example in which every row and value is invented, but the safe structure and the analysis task resemble the real problem.

real data stays in RCC
          |
          +--> invented coding example --> off-site coding agent
                                                |
                                                v
                                           general code
                                                |
                                                v
                              test and run inside RCC on real data

Bring the returned code into RCC, inspect it, test it on the invented example, and then run it against the real input inside RCC. Real inputs, revealing error messages, and real outputs do not go back to the off-site agent.

Class 18 shows how to make the synthetic coding example and bring the code back safely. It includes a downloadable example bundle and a prompt you can reuse. The documented manual method is available now; Make coding example is the intended future shortcut.

Which route should I choose?

Situation	Route
The agent runs inside RCC and is available for this project	A — work inside RCC
The agent is a website, desktop service, or other system outside RCC	B — synthetic data only
You cannot tell where the agent processes or stores prompts and files	Do not attach the real data; use B
A useful synthetic example cannot be made without exposing protected detail	Use A or continue without an agent

What not to send outside RCC

Do not paste or upload real project files, real rows, identifiers, free text, exact dates, revealing filenames, screenshots, real outputs, or unredacted error messages to an off-site coding agent. Shuffling rows, hashing identifiers, or changing names does not make the remaining data synthetic.

For the fuller data boundary, read before sharing data with a coding agent.

The goal is still simple

Whichever route you use, ask for code that:

accepts explicit input and output paths;
does not modify its input;
records its command, package versions, and logs;
is tested on small invented data first; and
uses RCC workers for substantial computation.