Roblox key art showing the Roblox logo surrounded by images from games on the platform. Image: Roblox
A web crawler that identified itself as Roblox’s copied 145.5 GB of artists’ 3D models from three free model-sharing sites, including one that hosts only adult content, the sites’ operators say. In a post on Monday, the team behind Open3DLab, SFMLab and SmutBase said the crawler ignored their robots.txt file and terms of service, and that they believe the files were gathered to train AI.
Roblox hasn’t commented. The sites say two emails to the company went unanswered.
What the logs show
According to the post, the crawler identified itself on every request as trellis2-data-sourcing/1.0 (3D foundation model research), followed by a Roblox work email address, which the sites have redacted. It came from a single Amazon Web Services address.
- Pages read: 18,926 project pages, almost the whole catalogue, in 26,429 requests over 34 hours.
- Files taken: on October 15 and 16, complete copies of 1,027 files from 712 projects, 145.5 GB in all.
- Licences: 423 of those 712 projects are licensed for non-commercial use only.
- Adult content: 224 of the projects came from SmutBase and 121 more were marked adult on SFMLab, 345 in total (62.6 GB).
The team says it spotted the crawler on October 18 and blocked it that morning. It came back on October 19 and 20, kept up its attempts in the face of the blocks, and worked what looked like office hours in central Europe, pausing around lunch. The sites say they have kept the full logs for any outside audit.
Why they point to Roblox’s 3D AI
Roblox unveiled Cube, which it calls its core generative AI system for 3D and 4D, in March 2025 and open-sourced a version of it. The accompanying paper says: “Our goal at Roblox is to build such a foundation model for 3D intelligence.” The sites say the email in the user agent belongs to one of the paper’s 40-plus authors, and stress that they hold Roblox, not the researcher, responsible.
“Trellis2” may refer to TRELLIS.2, an open 3D generation model Microsoft released in late 2025, or may just be a codename. The sites say its published data toolkit contains no crawler, so this one was written for the job. Their pitch for why someone would want their files: complete, textured, often rigged characters in the formats artists work in, which public datasets have little of.
They are careful about the limits. Anyone can type any email address into a user agent, and the traffic came from Amazon’s cloud rather than Roblox’s own network, though they say nothing suggests impersonation. They also say they can’t prove the data was for training, or that any of it reached a model.
“Nobody checked”
The adult material is what makes this awkward for Roblox. The company estimated in a 2025 filing that about 40% of its daily users in 2024 were under 13, and it tightened its rules on sexual content last year. The sites say they don’t think Roblox set out to collect adult content, but the crawler took it at the same rate as everything else:
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.
Salaryman, Open3DLab
They add that when a creator deletes a project, or the sites remove one after a copyright notice, it disappears from their pages but not from anyone’s dataset. One project the crawler took has already been disabled on SFMLab.
Roblox isn’t the only AI company in the logs. The team says Meta’s crawler, Meta-ExternalAgent, made about 380,000 requests in the month before and loaded 211,000 project pages, but Roblox’s was the first they spotted actually downloading users’ files. Meta’s crawling is already under fire from businesses over its new commercial terms, and the sites’ experience is another reminder that robots.txt alone won’t stop AI bots.
The sites have published six questions for Roblox, including whether it ran the crawler, what the data was for, whether anyone reviewed which sites it visited, and whether it will delete the files and anything derived from them.
Why it matters
Most fights over AI training data involve books, news and music, whose owners can afford to sue, as Canada’s big newspapers are doing with Anthropic. This one involves hobbyist 3D artists who chose non-commercial licences, and a company that builds generative tools for a platform full of children. If the logs are right, nobody screened what the crawler collected.
Sources: Open3DLab; Roblox, Introducing Roblox Cube; Cube: A Roblox View of 3D Intelligence (arXiv); Roblox 10-Q, September 2025; Roblox, policy on romantic and sexual content.
Latest More Labs news
