Crawler claiming Roblox took 145 GB of 3D models, adult content included
A crawler identifying itself as Roblox downloaded 145.5 GB from Open3DLab, SFMLab and SmutBase, including adult projects; the user-agent claim is unconfirmed and Roblox has not replied.
Open3DLab says a crawler identifying itself as Roblox scraped its three 3D-model sites on 15-16 October 2026, reading 18,926 project pages and downloading 1,027 files from 712 projects, 145.5 GB in total. The haul included 364 files (54.8 GB) from SmutBase, an adult-only site.
Across the network, 345 projects (62.6 GB) were marked adult. The site blocked the crawler on 18 October and emailed the address in its user agent; Roblox did not reply.
The crawler resumed on 19 October but was blocked, then tapered off on 20 October. Open3DLab says the user agent is a self-description and it cannot prove the files reached a model.
Key takeaways
- The crawler read 18,926 project pages in 26,429 requests over 34 hours, then downloaded 1,027 files from 712 projects. Free-download server caps on speed and concurrency limited the volume; 423 of the 712 projects were licensed for non-commercial use only.
- Adult content made up a large share: 364 files (54.8 GB) came from SmutBase, and 345 projects (62.6 GB) were marked adult. Open3DLab says the crawler took adult projects at the same rate as everything else and that one downloaded project has since been disabled on SFMLab.
- The crawler's later attempts fit office hours, pausing briefly around lunchtime and ending near the close of the business day in CEST. On 20 October it returned at 8:00 and tapered requests from lunchtime, stopping around 15:00, which Open3DLab reads as testing its rate limiting.
- The user agent, trellis2-data-sourcing/1.0, named a researcher among 40-plus authors of Roblox's Cube 3D foundation-model paper. Open3DLab speculates the data was for training or evaluation, possibly using Microsoft's TRELLIS.2 tooling, but says it cannot prove any files reached a model.
- Open3DLab also reports Meta-ExternalAgent made about 380,000 requests and loaded 211,000 project pages in the prior month, and it has blocked that crawler too. Open3DLab cites AI copyright cases involving Meta, Nvidia and Anthropic to argue enforcement is two-tiered.
What people think
Analysis details
26 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 26 of 39 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Cloud IP blocks cut bots but hit legitimate users
Blocking AWS, Azure and DigitalOcean ranges can suppress crawlers, but legitimate office VPNs and commercial VDI often egress from those networks. One reported AWS WAF rule blocked a customer's GCP-hosted VDI and was removed after causing harm without stopping bots. Bot traffic increasingly arrives through residential proxies, making datacenter blocks a partial filter; the disagreement is whether the lost access is acceptable.
Ripped assets undercut the moral complaint
Open3DLab's catalogue includes fan ports and ripped commercial game assets, some tagged with permissive licenses, and one video shows ripping encrypted Unreal Engine models. Critics call that theft and hypocrisy; defenders say Open3DLab admits the grey area, complies with takedowns and targets unequal enforcement, where individuals face ruinous damages and large AI firms treat fines as a cost of business.
Download alone does not prove license breach
A download from a free-access server does not itself violate license terms, and no supplied evidence shows the files reached a model. The counterpoint cites the crawler's retrieval of robots.txt, its disregard for terms of service and 423 non-commercial projects. If training is derivative and commercial, licenses such as CC-BY-ND-NC would be breached by use, not by the act of copying.
Training use differs from redistributing files
One argument separates training on data from using it one-to-one: current law may permit training even when redistribution would not. That distinction matters for the license question, because non-commercial and no-derivatives terms restrict both commercial deployment and derivative works. The disagreement is whether model training falls inside those restrictions or sits outside them until a court says otherwise.
Waiting for encryption to weaken
A separate proposal suggests delaying extraction of encrypted game assets until AES becomes easy to break, by which time the assets would be old enough to feel archival. The idea does not address current consent or license terms, and no technical timeline for breaking AES is supplied.