Roblox has confirmed that an AI research agent built by one of its researchers downloaded 145.5 GB of creators’ 3D models from three independent model-sharing sites — and says none of that data will be used to train its AI models.
The confirmation, reported on October 28, follows a disclosure two days earlier by Open3DLab, the community site that first spotted the crawler. According to Open3DLab’s account, the crawler downloaded 1,027 files from 712 projects across Open3DLab, SFMLab and SmutBase on October 15 and 16, 2026.
What the logs show
Open3DLab says the crawler identified itself on every request as “trellis2-data-sourcing/1.0 (3D foundation model research)” and carried a Roblox work email address. Over about 34 hours it made 26,429 requests from a single Amazon Web Services address, reading 18,926 project pages — nearly the entire catalogue of the three sites — before downloading complete copies of the files.
The sites say the crawler fetched their robots.txt files, the standard files that tell automated visitors what they may crawl, and then disregarded them. It also ignored the sites’ terms of service, which prohibit scraping. Of the 712 projects downloaded, 423 were licensed for non-commercial use only, according to the operators.
The adult-content question
345 of the downloaded projects were adult content: 224 from SmutBase, which hosts only adult material, and 121 more marked as adult on SFMLab. That detail is what turned a data-scraping dispute into a child-safety debate. Roblox is a platform used by tens of millions of children, and Open3DLab’s administrator, known as Salaryman, warned that if adult training data reached a model that generates assets for Roblox users, only the company’s filters would stand between that data and children.
Roblox’s statement does not concede that outcome. A company spokesperson told Game Developer and CyberInsider that the researcher-built agent was meant “to collect publicly available 3D assets” but “mistakenly” scraped Open3DLab and was “failing to operate as intended.”
“The agent was decommissioned, and the data was not used for any purpose, including training any AI models,” the spokesperson said. “We are strengthening safeguards for how agents like this are built and reviewed, and will continue to do so as the agentic landscape evolves.”
Why creators are angry
The three sites are small, independent communities, not commercial stock libraries. SFMLab launched in 2014 for models made with Valve’s Source Filmmaker software, SmutBase followed in 2016 for adult-only content, and Open3DLab opened in 2020 for general 3D work. Creators upload characters, props and environments for free and choose a licence for each project.
Open3DLab says it spotted the crawler on October 18, blocked it and emailed the address in the crawler’s identifier. The crawler resumed attempts on October 19 and 20, the site says, and only stopped after further blocks. The operators say they have kept full logs for any outside audit.
One figure in this story is not settled and must be presented with that caveat: different reports of the operators’ account give different totals for how much of the haul was adult content — 62.6 GB in some reports, 117.4 GB in Game Developer’s account. What all reports agree on is the overall total, 145.5 GB, and the count of 345 adult projects.
Why it matters beyond Roblox
The incident lands in the middle of a wider fight over how AI companies source training data. The crawler’s own identifier described its purpose as “3D foundation model research,” and Roblox already operates Cube 3D, an open-source AI model for generating 3D objects from text prompts. The operators suspect the scraped files were gathered for AI dataset training; Roblox has not confirmed that purpose, only that the data was not used.
For creators, the case is about consent and licences: a non-commercial licence is not permission to feed a commercial AI pipeline, and a robots.txt file is not a suggestion. For platforms, it is a warning about autonomous agents — software that can walk an entire catalogue in 34 hours needs checks before it is switched on, not apologies after.
What happens next
Open3DLab has not said whether it will pursue legal action. Roblox has not said whether the downloaded files have been deleted — the operators asked for deletion, and the company’s statement names only Open3DLab, not SFMLab or SmutBase. Its claim that the data went unused also cannot be checked independently. What is confirmed is narrower: the agent is shut down, Roblox says the data was never used, and the logs exist. The next test will be whether other AI labs treat this as a cautionary tale — or as a template that only failed because it was caught.
Frequently asked questions
Did Roblox use the scraped models to train AI? According to Roblox, no. The company says the data “was not used for any purpose, including training any AI models.” The operators suspected a training purpose because of the crawler’s identifier, but that purpose is not confirmed.
How much data was taken? 145.5 GB in total: 1,027 files from 712 projects, downloaded on October 15-16, 2026.
What is Open3DLab? An independent community site hosting free 3D models, alongside sister sites SFMLab and SmutBase. Creators choose licences for their uploads, and 423 of the downloaded projects allowed non-commercial use only.
What did Roblox say went wrong? That a researcher’s agent built to collect publicly available 3D assets “mistakenly” scraped Open3DLab and was “failing to operate as intended.” The agent has been decommissioned.
Sources
Open3DLab (primary, October 26, 2026): https://open3dlab.com/news/roblox-downloaded-145-gb-of-models-from-our-sites/ — Game Developer (October 28, 2026): https://www.gamedeveloper.com/business/roblox-confirms-no-adult-content-used-in-ai-training-after-scraping-open3dlab-data — RuntimeWire (Roblox statement, October 27, 2026): https://runtimewire.com/article/roblox-says-a-researcher-s-ai-agent-mistakenly-scraped-open3dlab — CyberInsider: https://cyberinsider.com/roblox-ai-scraper-3d-model-decommission/
Related on Emadan: Anthropic IPO prediction markets — https://emadan.store/anthropic-ipo-prediction-markets-betting-valuation-venue-october-28/ — Oracle, ChatGPT Codex and Google TPU cloud — https://emadan.store/oracle-chatgpt-codex-apps-google-tpu-cloud — GPT-6.1 Terra on Amazon Bedrock — https://emadan.store/gpt-6-1-terra-amazon-bedrock-generally-available
