Roblox downloaded 145 GB of models from our sites (including adult content) without asking
Pangram verdict · v3.3
We believe that this text is a mix of AI and human-written content.
AI likelihood · overall
MixedArticle text · 1,614 words · 25 segments analyzed
On 15 and 16 October 2026, a crawler that said it belonged to Roblox walked through almost every project on Open3DLab, SFMLab and SmutBase, and proceeded to download complete copies of 1,027 files from 712 projects. 145.5 GB of our creators' work, downloaded for what we assume are AI training purposes.This post briefly describes what happened, what mitigations were taken, and what this signals for the community at large. A more detailed writeup outlining what we saw in our logs will be posted at a later time.What happenedOn Sunday the 18th of October, we noticed a crawler making a large number of requests to our sites and download infrastructure. The crawler sent the same user agent on every request, trellis2-data-sourcing/1.0 (3D foundation model research; [redacted]@roblox.com). It came from a single Amazon Web Services (AWS) address. It had read the main page of 18,926 projects, which is practically our entire catalogue, in 26,429 requests over a span of 34 hours.After pulling the project pages, it downloaded complete copies of 1,027 files from 712 projects, 145.5 GB in all, before suspending operations during the weekend. The downloads include 364 files (54.8 GB) from 224 projects on SmutBase, which hosts adult content only. With the SFMLab projects marked adult, 345 of the 712 projects (62.6 GB) are adult content.The main limiting factor in the number of files downloaded was the crawler's reliance on the free download servers, which have caps on speed limit and number of concurrent downloads. The crawler did not respect our robots.txt, (which it did retrieve), ignored our Terms of Service, (which forbid scraping), and defied the creators' licenses: 423 of the 712 projects are licensed for non-commercial use only. Within the sample it retrieved, it took adult projects at the same rate as everything else. Some examples include models of Eve from Stellar Blade, Mei from Overwatch, and a Horse cock dildo.Our responseWe found it on 18 October and blocked further attempts the same morning. The scraper had suspended its activity by that point, and we did not notice any scraper that weekend. We emailed the address in its user agent on 18 October and asked Roblox for comment. We have not had a reply.The crawler resumed scraping activity on October 19th, with all requests getting blocked by our servers. The scraper did not reduce its download attempts when faced with our blocks, and the scrapes seemed to fit neatly within office hours, only briefly pausing for an hour around lunchtime, and ending activity around the end of the workday (CEST). On the 20th, their scraper made its final return at 8:00, slowly tapering requests from lunchtime onward, likely in an effort to test our rate limiting. The requests stopped around 15:00. We sent another email to Roblox for questions, to which they again did not reply.SpeculationThis section contains our speculation and assumptions based on public information, and the information we noticed in our logs.Roblox has been building a "3D foundation model" since at least 2025.In March 2025 Roblox introduced Cube, which it calls its core generative AI system for 3D and 4D.
The accompanying paper describes the goal as a foundation model for 3D intelligence.
It lists character rigging among the things that model should eventually do, and generating characters and avatars as future work. Its first shape model was trained on roughly 1.5 million 3D assets from licensed and public datasets such as Objaverse, plus free assets from Roblox's own Creator Store whose creators had opted in. In February 2026 Roblox began rolling out "4D generation" built on Cube.The address in the user agent belongs to a researcher who is among the 40-plus authors of that Cube paper. The researcher's public professional profile says they are building a 3D foundation model at Roblox, and their earlier published work is on 3D reconstruction and real-time 3D rendering. We want to be clear: we don't suggest this researcher did anything on their own initiative or out of the ordinary for their job.
A data-collection run like this is a company's decision, and the responsibility for it is Roblox's.
But the profile says what the user agent says: the people behind this crawler build 3D foundation models.The "trellis2" in trellis2-data-sourcing may refer to TRELLIS.2, an open 3D generation model Microsoft released in late 2025. It may mean the data was meant for training or tuning a model like it, or it may just be a project codename.
We don't know.
What we can say is that TRELLIS.2's published data toolkit contains no web crawler and no user agent, so this crawler and its label were written for the job. It is also possible that Roblox is re-using the Trellis.2 tooling for its own foundational model and it is re-using these scripts for its next model; the tools are open source.Our sites are full of things a 3D model-generation system needs and public datasets have relatively little of: complete, textured characters, many of them rigged, in the formats artists actually work in, each with a title, a description and preview images.Put together, we think the most likely explanation is that the crawler was collecting training or evaluation data for Roblox's 3D foundation model work.
We can't prove it, and we don't know whether any of our creators' files ended up in a model.Roblox's track recordThis isn't the first time Roblox has been criticised for how it treats the people whose work its platform runs on, or for how it handles other people's rights.Young developers.
In 2021 the games outlet People Make Games published a video investigation arguing that Roblox turns its young users into underpaid developers.
By its calculation Roblox kept around 75% of the money spent in developers' games, against Steam's 30%, and paid creators in Robux, which they can only cash out once they've earned a minimum amount. Roblox declined to discuss its revenue share and payout minimums.
(In September 2025 it raised its cash-out rate by 8.5%.)Copyright.
In 2021 the US National Music Publishers' Association sued Roblox for at least $200 million over unlicensed music on the platform.
The case settled later that year for an undisclosed amount.Child safety.
In October 2024 the short seller Hindenburg Research accused Roblox of inflating its user numbers and of failing to protect children.
Roblox called the report misleading and pointed out that its authors stood to profit from a falling share price.
Since then the attorneys general of Louisiana, Kentucky, Texas and Florida, among other states, have sued Roblox over child safety, and private lawsuits have been consolidated in US federal court. These are allegations, not findings.None of this proves anything about the crawler, but it strengthens our opinion that Roblox is a deeply unethical business.
It shows a company that has repeatedly been accused of putting growth ahead of the people its platform depends on.You see how this looks, right?More than a third of Roblox's daily users were under 13 in 2025, by its own figures.
Its rules ban sexual content, and in 2025 it tightened them further.Roblox's crawler took 224 projects from a site called SmutBase, which describes itself as exclusively for adult content, and 121 more that are marked adult on SFMLab. It took them at the same rate as everything else: 345 of the 712 projects it downloaded from are adult content.
The further scraping attempts did not discriminate between adult content either.We want to be careful here.
We don't think Roblox set out to collect adult content, and we have no evidence that any of it reached a model or a product.
The scrapes were likely automated. But that is exactly the problem: as far as our logs show, nobody checked. A crawler working for a company whose platform tens of millions of children use was pointed at an adult site, and it took explicit models without telling them apart from anything else.Models reproduce what they learn from. If adult training data reaches a model that generates assets for Roblox, the only thing between that data and children is whatever filter Roblox builds afterwards, because its data collection didn't filter anything.
Furthermore, when a creator deletes a project, or we remove one for breaking our rules or after a copyright notice, it's gone from our sites, but remains in these datasets. One project the crawler took has already been disabled on SFMLab since. If Roblox still has its copy, that copy is still there.We're not against adult content. It's our bread and butter. We think consenting adults should be able to access and enjoy this content freely, without the government looking over their shoulder. What we object to is our adult sites being emptied into an exploitative gaming company's training data without anyone asking, and being mixed into content offered to literal children.For a company being sued by several US states over child safety, "our AI team bulk-downloaded an adult site without looking" is not a good position to be in.If the penalty for a crime is a fine, then that law only exists for the lower class.This section is opinion.Roblox isn't the worst offender in AI's appetite for other people's work, just the latest to turn up in our logs.
It isn't the only AI company in them, either.
In the month before we found trellis2-data-sourcing, Meta's AI crawler, Meta-ExternalAgent, made about 380,000 requests to our sites, almost all from Meta's own network, and loaded 211,000 project pages. We have since blocked their crawler as well.The Roblox trellis2 crawler is the first individual crawler we spotted that actually pulled assets uploaded by our users, in flagrant defiance of any copyright or license that might apply to it.
It fits a broader context, of AI labs taking content without permission.