Connect with us

GAMING

Roblox Retired Its Cube Crawler After an Unfiltered Haul

Roblox shut a Cube-linked agent after it pulled 145.5 GB of 3D files, including adult models, with no content filter at collection.

Published

on

Roblox shut down a research agent on October 28 after it copied 145.5 GB of 3D files from Open3DLab, SFMLab and SmutBase. A spokesperson said the agent was built to collect public 3D assets, scraped those sites by mistake, and that the files will not train any model.

The logs describe a different failure. Adult projects came off the same queue as everything else, at the same rate, and nobody appears to have checked the pile before the operator did.

A Cube-Linked Agent Walked the Whole Catalogue

Open3DLab operator Salaryman wrote on October 26 that a crawler using the user agent trellis2-data-sourcing/1.0 (3D foundation model research, with a redacted @roblox.com address) walked through almost every project on Open3DLab, SFMLab and SmutBase. It ran from a single Amazon Web Services address. Over 34 hours on October 15 and 16, 2026, it read the main page of 18,926 projects in 26,429 requests, which Salaryman called practically the entire catalogue, then copied 1,027 files from 712 projects.

The agent retrieved the sites’ robots.txt file and did not follow it. It also ignored terms of service that forbid scraping. The free download servers, which cap speed and concurrent transfers, were what kept the file count from going higher. Among the sample it took were fan models of Eve from Shift Up’s Stellar Blade and Mei from Blizzard’s Overwatch, both hosted on SmutBase.

The user agent named a Roblox researcher who, Salaryman said, is among the 40-plus authors of the Cube paper on 3D intelligence. The person’s public profile describes building a 3D foundation model at Roblox. Open3DLab withheld the name, arguing the crawl was a company act. The trellis2 label may nod at Microsoft’s TRELLIS.2 3D model, or it may be an internal codename. Salaryman noted that TRELLIS.2’s published data toolkit contains no crawler and no user agent, so this one was written for the job.

THE CRAWL CLOCK

  1. October 15-16, 2026: Crawler reads 18,926 project pages and copies 1,027 files (145.5 GB), then goes quiet for the weekend.
  2. October 18, 2026: Open3DLab spots the traffic, blocks it that morning, and emails the address in the user agent. Roblox does not reply.
  3. October 19, 2026: Scraping resumes and is blocked. Attempts sit inside CEST office hours, pause around lunch, and fade at the end of the workday.
  4. October 20, 2026: Final return at 8:00. Requests taper after lunch and stop around 15:00. A second email to Roblox also goes unanswered.
  5. October 26, 2026: Salaryman publishes the logs and six questions.
  6. October 28, 2026: A Roblox spokesperson confirms a researcher built the agent, says it is shut down, and says the data was not used.

A shutdown after two ignored emails is not the same thing as a filter that ran while the agent was live. The October 19 and 20 retries also look like a scheduled workday job, not a bot that slipped its leash overnight.

345 Adult Projects, Taken at the Same Rate

SmutBase hosts adult content only. The crawler still took 364 files (54.8 GB) from 224 projects there. Counting SFMLab projects marked adult, 345 of the 712 projects, 62.6 GB, are adult. Salaryman wrote that adult projects were taken at the same rate as everything else, and that later blocked attempts did not discriminate either.

THE HAUL BY SITE

Slice Projects Files Size
SmutBase (adult only) 224 364 54.8 GB
Adult-marked across the three sites 345 n/a 62.6 GB
Full crawl 712 1,027 145.5 GB
Noncommercial licenses 423 n/a n/a

Salaryman was careful on intent. Open3DLab does not think Roblox set out to collect adult work, and it has no evidence any of it reached a model. The complaint is the gap that sentence leaves: as far as the 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.

Salaryman, Open3DLab operator, October 26, 2026 post

He added that models reproduce what they learn from. If adult files reach a model that generates assets for Roblox, the only barrier left is a filter built after collection, because collection did not sort anything. One project the crawler copied has already been disabled on SFMLab. If Roblox kept its copy, that copy is still there.

Why Cube Still Needs Other People’s Meshes

In March 2025, Anupam Singh, vice president of engineering, and Nick Tornow, vice president of creator engineering, presented Cube as Roblox’s core generative AI system for 3D and 4D. The company open-sourced a Cube 3D model and shipped mesh generation in Studio. Cube tokenizes shapes and predicts the next shape token the way a language model predicts the next word. In February 2026 it began rolling out 4D generation on top of that stack, so a generated car can be driven rather than sitting as a still mesh.

The Cube paper is blunt about the bottleneck. Readily available 3D training data is much thinner than the piles used for text, image or video models. The first shape model, Open3DLab notes, trained on roughly 1.5 million assets from licensed and public sets such as Objaverse, plus free Creator Store items whose makers had opted in. That is a closed, described diet. A public crawl of someone else’s catalogue is not.

CUBE’S PUBLISHED TRAINING PILES

  • First shape model: About 1.5 million 3D assets from licensed and public sets plus opted-in Creator Store files.
  • CubePart stage 1: Trained on roughly 4.7 million mesh-text pairs, which Tinghui Zhou called the data-hungry stage.
  • CubePart dataset: More than 460,000 assets and 2.02 million parts, over 11 times larger than a prior public set named in the post.
  • Stated aim: A foundation model that can make objects, scenes, rigged characters and scripts for Roblox experiences.

Zhou’s May 28, 2026 CubePart note says mapping open-vocabulary language onto 3D geometry needs a large, diverse corpus to do well. Salaryman made the matching point from the other side of the firewall. Open3DLab’s sites are full of complete, textured characters, many of them rigged, in the formats artists actually work in, each with a title, a description and preview images, the sort of material public datasets still lack. Put together, he wrote, the most likely explanation is that the crawler was collecting training or evaluation data for Roblox’s 3D foundation model work. He also wrote that he cannot prove it, and does not know whether any file landed in a model.

Roblox Shut the Agent and Left Files Hanging

Two days after the public post, Roblox confirmed the researcher and the malfunction. It did not answer the rest of the list Open3DLab had put on the record.

A Roblox researcher built an agent to collect publicly available 3D assets that mistakenly scraped content from Open3DLab, failing to operate as intended. The agent was decommissioned, and the data was not used for any purpose, including training any AI models. We are strengthening safeguards for how agents like this are built and reviewed, and will continue to do so as the agentic landscape evolves.

Roblox spokesperson, statement on October 28, 2026

That statement covers ownership, a claim of no training use, and a future review process. It does not say why robots.txt was fetched and then ignored, whether anyone reviewed target sites, whether adult catalogues were excluded at any stage, whether the files and anything derived from them have been deleted, or whether noncommercial licenses will be honored on the next run. Calling a 34-hour, full-catalogue copy a mistake also sits beside a crawler that came back for two workdays after it was blocked.

SIX QUESTIONS OPEN3DLAB PUT TO ROBLOX

  • Who ran it: Did Roblox run the trellis2-data-sourcing crawler? The company now says a researcher built the agent.
  • The purpose: What was the data for, and has any of it been used to train or evaluate a model? Roblox says it was not used for any purpose, including training.
  • The rules: Why did the crawler ignore robots.txt and the terms of service? Unanswered.
  • The review: Did anyone review which sites it would visit, and was adult content excluded at any stage? Unanswered.
  • Deletion: Will Roblox delete the files and anything derived from them, and confirm that in writing? Unanswered.
  • Licenses: Will it respect creators’ noncommercial licenses from now on? Unanswered.

Respecting robots.txt is one check before a download. The agent had already retrieved the file. An agent can fail at targeting and still be very good at copying, which this one was. Shutting it down after the operator published the logs is the easy half of the cleanup. Saying the 145.5 GB is gone, in writing, is the half that is still missing.

Noncommercial Licenses Did Not Follow the Meshes

Salaryman wrote that 423 of the 712 projects are licensed for noncommercial use only, and that the crawler defied those terms. Many of those licenses sit on models based on other companies’ characters, which is why Eve and Mei show up in a research crawl at all. A fan mesh on SmutBase is not a license for a public company to vacuum the file into a foundation-model pile.

Open3DLab is not shy about adult work. Salaryman called it the sites’ bread and butter, and said consenting adults should be able to use it without the government looking over their shoulder. The objection is the mix: adult catalogues emptied into a gaming company’s training data without anyone asking, then pointed at a platform used by children.

The same month, Open3DLab said, Meta’s AI crawler Meta-ExternalAgent made about 380,000 requests and loaded 211,000 project pages, almost all from Meta’s own network. That crawler was blocked too. Salaryman said the Roblox agent was the first individual crawler the logs caught that actually pulled assets users had uploaded. Hitting a page is one kind of scrape. Walking off with the files is another.

What the Logs Show About the Next Agent

Roblox’s own rules ban sexual content, and in 2025 the company tightened its policy on romantic and sexual material. Salaryman noted that more than a third of daily users were under 13 in 2025, by Roblox’s figures, and that the platform is used by tens of millions of children. A company in that position is the one that can least afford an intake pipe with no adult filter on it.

That risk sits on top of lawsuits it is already fighting. Open3DLab pointed to attorneys general in Louisiana, Kentucky, Texas and Florida, among other states, suing over child safety, plus private cases consolidated in federal court. Those are allegations, not findings. Roblox has denied the claims. They still make “our AI team bulk-downloaded an adult site without looking,” as Salaryman put it, a bad place to stand.

The October 28 line about strengthening how agents are built and reviewed is a promise about the next crawler, not a description of this one. This one fetched robots.txt, ignored it, treated SmutBase like any other public 3D host, and kept knocking after it was blocked. Cube’s published work still says 3D data is scarce and that the next stages are hungry. Other people’s rigged, textured characters remain the shortest path to a fatter corpus.

Salaryman asked Roblox to delete the files and anything derived from them, and to confirm that in writing. The October 28 statement does not say they did.

Harry is the editor and lead writer of STUDIO ONE NETWORKS, an independent title he owns and runs himself. Ten years in journalism, reporting first and editing later, taught him that entertainment and business are one beat seen from two sides: a box office figure is a company number, a streaming deal is a contract, a casting rumour is not a story until someone puts their name to it. He works from the record, whether that is a distributor's statement, a licensing agreement, an interview transcript or a set of published ratings, and checks every number against it before publication. The same rule holds for the rest of the site, which covers news, technology, science, sports, lifestyle, travel, auto and gaming for an audience spread across the world. When he gets something wrong, the article is corrected and the change is noted and dated, under a corrections policy anyone can read. Reader mail is answered by him at support@studioonenetworks.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending