Open3DLab Reports Roblox-Linked Crawler Downloaded 145.5GB of 3D Assets

| 2 Min Read
Open3DLab says a crawler using a Roblox-linked research identifier scanned nearly its full catalog and downloaded 1,027 files, including adult-marked assets.

Open3DLab says a crawler identifying itself with a Roblox-linked 3D research user agent downloaded 145.5GB of assets from its sites, including material marked for adults. The community’s account raises a practical question for teams building 3D foundation models: how do they verify that scraped training data respects creator permissions and content labels?

A broad scan, followed by file downloads

In a blog post, Open3DLab reports that the crawler mapped 18,926 projects across Open3DLab, SFMLab and SmutBase in 34 hours, making 26,429 requests. It then obtained complete copies of 1,027 files from 712 projects before suspending operations.

The activity began on October 18, when staff noticed a high volume of requests to the sites and their download infrastructure. The crawler’s user agent was trellis2-data-sourcing/1.0 (3D foundation model research; [redacted]@roblox.com), and requests came from a single Amazon AWS address. Open3DLab says the crawler ignored its robots.txt instructions and Terms of Service, both of which prohibit scraping. The site also says the downloaded projects carried creator licenses restricting use to non-commercial purposes.

The files included 364 assets totaling 54.8GB from 224 SmutBase projects. Another 345 adult-marked projects on SFMLab accounted for 62.6GB. Open3DLab says free download servers, which limit speed and concurrent transfers, were the crawler’s main constraint.

Blocking the crawler didn’t stop its attempts

Open3DLab detected the scraping on October 18 and blocked the crawler by the following morning. Its logs show more download attempts on October 19, which the site says continued at roughly office-hour intervals, with a short pause around lunch. Blocking the source did not reduce the attempts, according to the post.

Requests tapered on October 20, then stopped later that day. Open3DLab says the final activity may have been an effort to test its rate limits. It sought comment again at that point but says it received no reply.

The site has since made changes intended to reduce the chance of a repeat. Its account doesn’t detail those measures, but the incident shows how quickly a crawler can inventory a large catalog before a site can respond.

The model link is plausible, but unproven

The user agent’s reference to “3D foundation model research” is the basis for Open3DLab’s suspicion that the collection was meant for training or evaluation. Roblox introduced Roblox Cube in March 2025 as a generative AI system for 3D modeling and character rigging, with characters and avatars identified as future goals. In February 2026, the company launched a beta of “4D generation,” technology designed to create in-game models from natural-language prompts.

Open3DLab also says the email in the crawler’s user agent matches a researcher named on a Roblox Cube paper, though the site redacted the address. It calls data collection for the model its most likely explanation, while acknowledging that its logs can’t establish the crawler’s purpose.

The concern is sharper because the sites include adult content. Open3DLab says it doesn’t believe Roblox set out to collect those assets, but argues that its records suggest no one checked what the crawler was downloading. Roblox is a gaming and creation platform with more than 33 million users; about 40% are reportedly under 12. For model builders, the episode makes data filtering and permission checks concrete operational requirements, especially when a crawler can reach material creators explicitly marked as restricted.

Source: Ben Stockton · www.tomshardware.com

Comments

Please sign in to comment.