Roblox downloaded 145 GB of models from our sites (including adult content) without asking

extract

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 happened On 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 response We 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. Speculation This 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 record This 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. Here's what the largest companies have been accused of, and in some cases have admitted: - Meta downloaded at least 81.7 TB of pirated books by torrent from shadow libraries, including through Anna's Archive, according to court filings by the authors suing it. In June 2025 a judge ruled that training on the books was fair use on that record, while stressing that the ruling didn't make Meta's conduct lawful in general. - Meta, again: the adult studio Strike 3 Holdings alleges that Meta torrented thousands of its films to train AI video models. Meta says the downloads were for individuals' "personal use". In June 2026 the court refused to dismiss the case. The irony is hard to miss: Strike 3 is best known for suing thousands of individual BitTorrent users. - Nvidia contacted Anna's Archive about high-speed access to its collection of pirated books for training data, and management signed off soon after, according to an amended class-action complaint filed in January 2026. Nvidia argued that contacting a pirate site isn't infringement. The court has since largely refused to dismiss the case. - Anthropic downloaded hundreds of thousands of pirated books. A judge found that training on books it had bought was fair use, but keeping a library of pirated copies was not. In July 2026 the court approved a $1.5 billion settlement, about $3,000 per book. Now compare how copyright treats everyone else. In the early 2000s Jammie Thomas-Rasset was ordered to pay $222,000 for sharing 24 songs. Internet activist Aaron Swartz faced federal charges carrying decades in prison for downloading academic articles. The Internet Archive, a nonprofit library, lost in court for lending scanned copies of books it owned. Meanwhile, the Internet Archive's online library gets repeatedly scanned by AI companies in return. We'll be honest about where we stand. We don't like copyright as it exists. We'd like to see it radically reformed, and we wouldn't mourn it if it were abolished. Our community lives in its grey areas every day: many of the models on our sites are fan ports of characters from commercial games, made out of love and shared for free. When a studio sends us a takedown notice, we comply. While we as small artists and fan communities get copyright used against us, the world's largest companies are committing copyright infringement at a scale unseen in human history. This has been the case since the invention of the internet of course; digital media make duplication, remixing and resharing trivial, but copyright law never caught up. Filesharing stabbed copyright in the gut. AI is just twisting the knife, and completes the creation an unequal system where copyright infringement is only a crime if you're a low net worth individual. So we aren't asking for stricter copyright laws, or stronger protection for copyright holders. That will only benefit the major entertainment companies like Disney and Adobe who have already stacked the deck in their favor, and are happily making deals with Big-Tech and crushing small creators in the process. We're pointing out that the copyright we have isn't applied equally, and is in practice a two tier system. Individuals face ruinous damages for minor infractions. The largest companies in the world face a cost to be negotiated after the fact, if anyone can afford to sue them. A law that only works downwards gives nobody a reason to respect it, and it can't last in this form. For time being, we'd like to ask Roblox to respect the norms of this community. You credit people. You respect the license they chose. If you want to use someone's work for something they didn't offer it for, you ask. Those norms cost nothing to follow, and Roblox's crawler followed none of them. This scrape attempt saturated the connections for our free download servers for a couple of days, causing problems for our community of legitimate human users. It shows a lack of respect for the people who make this site what it is. those who produced, shared and host the work that you were about to throw in your AI mixer. Questions for Roblox We're sad that the emails we sent to Roblox were not met with a timely response, and would thus like to publicly ask Roblox for its response. These are the questions we'd like Roblox to answer: - Did Roblox run the trellis2-data-sourcing crawler? - What was the data for? Has any of it been used to train or evaluate a model, and if so, which? - Why did the crawler ignore our robots.txt rules and our Terms of Service? - Did anyone review which sites the crawler would visit? Was adult content excluded at any stage? - Will Roblox delete the files and anything derived from them, and confirm that in writing? - Will Roblox respect our creators' license choices, NonCommercial in particular, from now on? We hope Roblox reaches out to us. Also, if you are a person with information on what Roblox is doing, or a fellow site owner who has been targeted, feel free to get in touch as well. We're eager to hear if there's more of a pattern to the scraper's behavior. We've kept the full export of the relevant log files for auditing by external parties if the need arises. About us Open3DLab, SFMLab and SmutBase are three community sites for free 3D resources, run by a small team. SFMLab started in 2014 for Source Filmmaker models. SmutBase followed in 2016 and is exclusively for adult content. Open3DLab, for general-purpose 3D work, joined in 2020. Creators upload characters, props, maps and environments, choose a license for each project, and anyone can download the files for free. All three sites are intended for adults only. About the redaction. The email address in the user agent is a Roblox employee's work address. We've removed the name: this article is about what the Roblox crawler did, not about one person. About Attribution. We take the user agent at face value. Anyone can type any email address into a user agent, and the crawler ran from Amazon's cloud, not from Roblox's own network. Nothing we've seen suggests impersonation, and the label fits Roblox's published research closely. Still, it is a self-description, and it's the first thing we'd like Roblox to confirm.