Operate

Safety & moderation

What we accept, what we don't, and exactly which checks run today.

Read this section literally

Safety claims are the ones people act on, so this page describes only checks that actually run. Where a control is a policy we enforce reactively rather than a system that scans for you, it says so. An earlier version of this page described a red-team programme, malware and watermark scanning, license-conflict detection, and public per-listing safety reports. None of those exist. Assume any artifact you download is untrusted code and sandbox it.

Acceptable use

Node Data exists to move the field of physical AI forward. Assets are accepted across a wide range of use cases. The following categories are not allowed and are removed when we find them or when they are reported:

  • Models or datasets designed to identify, surveil, or target individuals without consent
  • Models trained on data obtained in violation of applicable privacy law
  • Assets whose primary purpose is to autonomously deliver kinetic harm
  • Generative models tuned to produce sexual content involving minors (CSAM)
  • Malware, including weights with embedded code execution payloads

This is a policy, and enforcement is human and largely reactive. Nothing in the upload path detects most of these automatically — see below for what is actually machine-checked.

What runs on upload

Every uploaded file goes through format validation. It confirms the file is structurally what it claims to be, and rejects it otherwise:

  • ONNX — protobuf header present
  • PyTorch — zip or pickle magic bytes, minimum size
  • JSON / YAML — parses, non-empty
  • NPZ — valid zip-based numpy archive
  • URDF / SDF — expected root element present
  • CSV — header plus at least one data row
  • ROS packages — archive integrity and a package.xml
  • Zip-family archives — readable central directory

That is a correctness check, not a security check. It tells you a .onnx file is really an ONNX file. It does not tell you the model is safe, and it does not inspect what a pickle would execute on load.

Not implemented

There is no antivirus scan, no hash blocklist of known-unsafe weights, no secret or embedded-URL scanning, and no license-compatibility checker. Uploads publish immediately; nothing automatically routes them to manual review.

Automated policy screen on AI Factory assets

Assets generated by AI Factory — and only those — get one additional check before they can be published. It is a language-model policy screen over the prompt, the analysis, and the generated artifacts, and it looks for:

  • Surveillance use
  • Kinetic harm
  • Biometric identification of individuals without consent
  • CSAM
  • Malware and embedded secrets
  • License conflicts

It returns a verdict of pass, review, or block, with categorised flags and a note, and defaults to review when uncertain. A block verdict does prevent publishing — that is enforced in code, not by convention.

The report is visible to the asset's creator in the Factory review UI and to moderators. It is not attached to the public listing page, and it does not travel with the asset when it publishes.

What this screen is not

It is a classifier reading text, not a red team. It does not run the model, probe it with adversarial prompts, test for prompt injection, or measure bias. Manually uploaded assets never see it at all.

Red-team review

Not available

There is no red-team programme, no capability threshold (by FLOPs, parameter count, or self-declared capability), no standard eval suite, and no public red-team report on any listing page. If you need adversarial assurance for a model you are deploying, you have to do it yourself.

Moderation queue

Moderators can see assets marked for review and can unpublish or remove a listing. Unpublishing removes it from the marketplace; existing buyers keep their download rights, so it is not a recall — if an asset is actively dangerous, the buyers have to be contacted directly.

Reporting abuse

Email trust@nodedata.ai with the listing URL, the category of concern, and any supporting evidence. There is no in-product “Report” button yet, and no automated intake — a report is a message a person reads, so please make it specific.

We are not going to publish a response-time commitment we have no system to measure. Urgent reports should say so in the subject line.

CSAM

CSAM is reported to NCMEC and the relevant authorities. Affected accounts are permanently banned and forfeit pending payouts.

Takedowns and appeals

When a listing is removed the seller is told why and can appeal by replying. Appeals are read by a person.

DMCA

DMCA notices go to dmca@nodedata.ai. We follow the standard notice-and-counter-notice procedure.