Quick Answer: In mid-October 2026, a crawler identifying itself as belonging to Roblox downloaded 1,027 files from 712 projects on Open3DLab, SFMLab and SmutBase, totaling 145.5 GB. The operator of the sites says the crawler ignored robots.txt and Terms of Service, took 345 adult projects at the same rate as everything else, and probably collected data for a 3D AI model. Roblox has not responded publicly, and the training purpose is unconfirmed.
Roblox 3D model scraping is not a hypothetical risk anymore. Over two days in mid-October, a single crawler read the project pages of nearly every model on three creator sites. It then pulled complete copies of hundreds of them. The operator of those sites, Open3DLab, published its account on 26 October, and it raises questions every creator who uploads 3D work should be asking. Here's what the logs show, what's still speculation, and what you can do about it now.
What the Crawler Actually Did
According to the Open3DLab post, the crawler sent the same user agent on every request: trellis2-data-sourcing/1.0 (3D foundation model research; ...@roblox.com). It came from a single Amazon Web Services address. Over 34 hours it made 26,429 requests and read the main page of 18,926 projects. That is effectively the entire catalogue.
Reading pages was only the first step. The crawler then downloaded complete copies of 1,027 files from 712 projects. The main limit was the free download servers, which cap speed and concurrent downloads. Without those caps, the volume would likely have been larger.
Here's the counter-intuitive part. Most people assume a scraper that lists a contact email is a responsible one. In this case, the contact address existed, but the crawler still ignored the site's robots.txt, which it had retrieved, and the Terms of Service that prohibit scraping. A declared identity is not the same as permission.
Why the User Agent Is the Interesting Clue
The user agent is the only direct evidence of purpose, and it's worth reading closely. The label says "3D foundation model research," and the Open3DLab team says the address belongs to a researcher listed as an author on Roblox's Cube paper. Roblox introduced Cube in March 2025 as its core generative AI system for 3D and 4D content. The paper describes character rigging and avatar generation as future goals.
That doesn't prove what the data was for. Open3DLab is careful about this, and so should you be. The string "trellis2" may point to TRELLIS.2, an open 3D generation model Microsoft released in late 2025. It might also be a codename. The site notes that TRELLIS.2's public toolkit contains no web crawler, which suggests this crawler was written for this job.
What makes the sites attractive to a 3D model team is the structure of the content. Creator uploads tend to be complete, textured characters, often rigged, in the formats artists actually use. Each comes with a title, description and preview images. Public 3D datasets such as Objaverse have plenty of objects, but far fewer complete, labeled characters. That's the gap this kind of data fills.
The Licensing Problem Most Creators Miss
Here's where most people stop reading, and where the real exposure lies. A file's license is not a formality. Of the 712 projects, 423 were licensed for non-commercial use only. Training a commercial model on those files is a use the license doesn't cover, regardless of whether the crawler had a good reason.
Many creators assume that uploading a model to a public site means they've granted broad reuse rights. That's a dangerous assumption. A license is a set of permissions, and it only binds people who respect it. Once a copy sits in someone else's dataset, you have no practical way to recall it.
This is also where the right to delete breaks down. When a creator removes a project, it disappears from the site. Copies pulled earlier stay in whatever dataset captured them. The source notes that one project the crawler took has since been disabled on SFMLab, but any copy Roblox holds is still there.
The Adult Content Issue Is Not a Side Note
The most uncomfortable detail is the mix. The crawler took 224 projects from SmutBase, which hosts adult content only. Another 121 projects on SFMLab are marked adult. In total, 345 of the 712 downloaded projects, about 62.6 GB, are adult.
Open3DLab's point is that the crawler did not sort content by rating. Within its sample, adult projects were taken at the same rate as everything else. Examples in the post include models of characters from Stellar Blade and Overwatch, plus explicit assets.
Think about why that matters. A model trained on mixed data can reproduce what it learned. Roblox's rules prohibit sexual content, and the company has tightened them. If adult material reaches a model that generates assets for a platform with a large under-13 audience, the only protection left is whatever filtering happens afterwards. Open3DLab's argument is that the filtering should have started at collection, and the logs suggest it didn't.
We don't know that any of this reached a model or a product. Treat it as a risk to be verified, not a confirmed outcome.
What You Can Do Right Now
You can't pull data out of a training set you don't control. You can reduce future exposure, and the steps below are practical and cheap.
- Check your license before you upload. Use a non-commercial or no-AI-training license if that's what you want. Note that license text alone is not enforced by any crawler.
- Publish a clear robots.txt and terms. Name known AI crawlers and state your scraping and training rules. Well-behaved bots respect this. Badly behaved ones don't, which is the point of the next step.
- Enforce at the server. Rate limits and user-agent blocks stop the easy cases. Open3DLab said its blocks were ignored at first, so pair them with logs you actually review.
- Keep your own records. Save your original files and upload dates. If a dispute arises, you'll need proof of authorship and of the license you attached.
Here's the trade-off to keep in mind. Blocking AI crawlers protects your files from future collection, but it also makes your work harder to find for legitimate users and search engines. Most creators will need to decide how much visibility they're willing to trade for control.
Frequently Asked Questions
What is the Roblox 3D model scraping incident?
In October 2026, a crawler using the user agent trellis2-data-sourcing/1.0 downloaded 1,027 files from 712 projects on Open3DLab, SFMLab and SmutBase, about 145.5 GB. The sites' operator says the crawler ignored robots.txt and Terms of Service. Roblox has not publicly replied, and the purpose of the downloads has not been confirmed.
How do I stop an AI crawler from downloading my 3D models?
Start with a robots.txt that names known AI crawlers and disallows them. Add rate limits and user-agent blocks at the server, since some crawlers ignore robots.txt. Review access logs regularly. No method stops every scraper, so keep original files and upload records in case you need to prove ownership later.
Why does a model's license matter for AI training?
A license states what others may do with your work. Training a commercial model on a non-commercial file goes beyond what that license allows. The license doesn't stop a crawler by itself, but it gives you a basis for a complaint or takedown request, and it sets clear expectations for anyone who accesses your files.
Scraping like this will keep happening. Pick one step from the list above and do it this week: review the license on your oldest upload and update it if it's unclear. Then pass this to any creator who uploads 3D work and wonders whether their files are already in someone else's dataset.