On October 15-16, 2026, a crawler identifying itself as trellis2-data-sourcing/1.0 and linked to a named Roblox researcher systematically scraped Open3DLab, SFMLab, and SmutBase. Over 34 hours it read 18,926 project pages — nearly the entire catalogue — and downloaded 1,027 files from 712 projects totaling 145.5 GB. It ignored the sites' robots.txt, violated their Terms of Service which explicitly forbid scraping, and disregarded individual creator licenses: 423 of the 712 projects were licensed for non-commercial use only. The composition of the haul is the sharpest detail. Of the 712 projects downloaded, 345 contained adult content — 224 from SmutBase, which hosts exclusively adult material, plus 121 adult-flagged projects from SFMLab. That's 62.6 GB of explicit 3D models ingested at the same rate as everything else, with no filtering whatsoever. The crawler took models of a horse cock dildo alongside Eve from Stellar Blade without distinguishing between them. The user-agent string points directly to Roblox's 3D foundation model program. The email address belongs to a co-author of the Cube paper, Roblox's generative AI system for 3D and 4D content announced in March 2025 and trained on roughly 1.5 million 3D assets. The sites' operators are careful to note this was a company decision, not individual initiative. But the connection between the crawler's stated purpose — "3D foundation model research" — and Roblox's publicly documented AI ambitions is direct. The child-safety contradiction is extraordinary. More than a third of Roblox's daily active users were under 13 in 2025 by the company's own figures. The platform bans sexual content and tightened those rules in 2025. Roblox is currently being sued by attorneys general in Louisiana, Kentucky, Texas, and Florida over child safety. Into this context, its AI team pointed an unfiltered crawler at a site literally called SmutBase and downloaded everything. The operators' point is precise: models reproduce what they train on, and if adult data reaches a generative model serving children, the only barrier is a post-hoc filter — because the data collection filtered nothing. Roblox's silence amplifies the problem. The site operators emailed the address in the user-agent on October 18 and sent a follow-up. No reply came. The crawler resumed activity on October 19 after being blocked, continued testing rate limits on the 20th during what appeared to be CEST business hours, and only stopped after its requests were consistently rejected. The pattern — scrape first, ignore objections, test defenses — is not the behavior of a company that made an innocent mistake. This is not an isolated incident in the broader AI ecosystem. In the month before Roblox's crawler appeared, Meta's Meta-ExternalAgent made approximately 380,000 requests to the same sites and loaded 211,000 project pages. The difference is that the Roblox crawler was the first to actually pull complete asset files. The site operators frame this within the documented pattern of AI companies taking training data without permission — Meta's admitted use of pirated books, OpenAI's acknowledged scraping of copyrighted material, and the industry-wide assumption that taking first and litigating later is cheaper than licensing. The structural dynamic is plain: Roblox's 3D foundation model needs exactly what these creator communities produce — complete, textured, rigged characters with metadata — and public datasets lack. Rather than license the work, pay creators, or even ask permission, it sent a bot. When caught, it went silent. The creators whose work was taken have no practical recourse beyond blocking future scrapes. The data already downloaded is gone.
