Roblox Accused of Scraping 145.5GB of 3D Models

A web crawler that identified itself as a Roblox 3D AI research tool downloaded 145.5 GB of community-made 3D models from Open3DLab, SFMLab, and SmutBase on 15 and 16 October 2026, including 62.6 GB of adult content, the sites’ operators said on 26 October.

The crawler ignored the sites’ robots.txt file and their terms of service, which prohibit scraping. Roblox has not answered two emails asking whether it ran the operation. The crawler sent the same user agent string with every request: trellis2-data-sourcing/1.0 (3D foundation model research; [redacted]@roblox.com).

A user agent is the label software attaches to each web request to identify itself. Open3DLab’s account says the email address in this one belongs to a Roblox researcher. The operators withheld the person’s name, saying the incident concerns the company’s crawler rather than any one employee.

Over 34 hours, the bot made 26,429 requests from a single Amazon Web Services (AWS) IP address. It read the main page of 18,926 projects, which is practically the entire catalogue of the three sites. It then downloaded complete copies of 1,027 files from 712 projects before going quiet over the weekend.

According to the operators, the haul was capped mainly because the crawler used the sites’ free download servers, which limit speed and the number of simultaneous connections.

The three sites are run by a small team and host free 3D characters, props, and environments. SFMLab launched in 2014 for Source Filmmaker models, SmutBase followed in 2016 for adult-only content, and Open3DLab opened in 2020 for general 3D work.

Creators choose a license for each project. Of the 712 projects the crawler downloaded, 423 were licensed for non-commercial use only. The bot also fetched each site’s robots.txt, the standard file that tells bots what they may crawl, and then disregarded it.

How the Sites Detected and Blocked the Roblox-Labelled Crawler

Open3DLab spotted the traffic on Sunday, 18 October. It blocked the crawler that morning and emailed the address in the user agent, asking Roblox for comment. The bot returned on Monday, 19 October, and kept trying to download even though every request was refused. Its activity followed a Central European Summer Time (CEST) working day: it paused for about an hour around lunch and stopped near the end of business hours.

On 20 October, the crawler came back at 08:00. From lunchtime onward, it gradually reduced its requests, which the operators believe was an attempt to test their rate limits, and it stopped around 15:00. A second email to Roblox also went unanswered. Open3DLab says the scraping saturated its free download servers for a couple of days and slowed downloads for regular users. It has kept a full export of the relevant logs in case an outside party needs to audit them.

Why Open3DLab Links the Scrape to Roblox’s 3D AI Model

Roblox has been building a 3D foundation model since at least March 2025, when it introduced Cube as its core generative AI system for 3D and 4D. The Cube technical paper lists character rigging among the model’s goals and avatar generation as future work.

According to the paper, Cube’s first shape model was trained on roughly 1.5 million assets. They came from licensed and public datasets such as Objaverse, as well as free Roblox Creator Store assets whose creators had opted in. In February 2026, Roblox began rolling out 4D generation built on the Cube Foundation Model, which generates objects that move and work from text prompts.

Open3DLab says the researcher whose address appeared in the user agent is one of the paper’s more than 40 authors, and that the researcher’s public profile describes building a 3D foundation model at Roblox.

The “trellis2” label may refer to TRELLIS.2, an open 3D generation model Microsoft released in late 2025. The operators note, however, that TRELLIS.2’s public data toolkit contains no web crawler, so this bot was written for the job. Sites like these hold what public 3D datasets largely lack: complete, textured, and often rigged characters in the formats artists use, each with a title, description, and preview images.

The operators say the most likely explanation is that the crawler was collecting training or evaluation data for Roblox’s 3D model work. They stress that they cannot prove this and do not know whether any files ended up in a model.

They also point out a limit of their own attribution: the crawler ran from Amazon’s cloud rather than Roblox’s network, and “anyone can type any email address into a user agent.” Still, the operators say nothing in their logs suggests impersonation, and the label closely matches Roblox’s published research.

Adult content taken for a platform built around children

SmutBase, which hosts only adult material, lost 364 files totalling 54.8 GB from 224 projects. Another 121 of the downloaded projects were SFMLab projects marked adult. Together, 345 of the 712 projects, or 62.6 GB, are adult content, including fan-made models of Eve from Stellar Blade and Mei from Overwatch. The crawler took adult projects at the same rate as everything else, which the operators say shows no one filtered its target list.

That sits awkwardly with Roblox’s own audience. By the company’s figures, more than a third of its daily users were under 13 in 2025, and it tightened its ban on sexual content that year. Roblox also faces child-safety lawsuits from the attorneys general of Louisiana, Kentucky, Texas and Florida, among other states, though none of those allegations has been proven in court.

The operators say they have no evidence that any adult material reached a Roblox model or product. They also note that projects deleted from their sites survive in scraped copies; one project the crawler took has since been disabled on SFMLab.

Roblox is not the only AI company in the sites’ logs. In the month before the discovery, Meta’s AI crawler, Meta-ExternalAgent, made about 380,000 requests and loaded 211,000 project pages, mostly from Meta’s own network. The sites have since blocked it as well. Open3DLab says the Roblox-labelled bot was the first crawler it saw downloading users’ uploaded files, rather than only reading pages.

Open3DLab has published six questions for Roblox:

• whether it ran the trellis2-data-sourcing crawler;
• what the data was for, and whether it was used to train or evaluate any model;
• why the crawler ignored robots.txt and the terms of service;
• whether anyone reviewed the target sites or excluded adult content;
• whether Roblox will delete the files and anything derived from them, and confirm it in writing;
• whether it will respect creators’ non-commercial licenses from now on.

As of 27 October 2026, Roblox had not publicly responded or confirmed that the crawler was its own.