Roblox AI Scraping Claim: A Crawler Took 145GB of Creators’ 3D Models Without Asking
A Roblox AI scraping incident is the latest fight over AI training data. On October 26, 2026, the community site Open3DLab said a crawler carrying a Roblox employee’s email address had downloaded 145.5 GB of creators’ 3D models from its sites on October 15 and 16. The crawler ignored robots.txt, and close to half of the projects it took were adult content. Roblox had not replied to two emails at the time of publication.
This article is based on Open3DLab’s published report and the Hacker News discussion that followed. We have not seen the server logs ourselves.

Key Takeaways
- When: downloads on October 15–16, 2026; found on October 18; blocked attempts continued until October 20.
- How much: 1,027 files from 712 projects, 145.5 GB in total.
- What it ignored: robots.txt, the sites’ Terms of Service, and creator licenses. 423 of the 712 projects allow non-commercial use only.
- Adult content: 345 of the 712 projects (62.6 GB) were adult content.
- Who: the user agent said
trellis2-data-sourcing/1.0 (3D foundation model research)with a @roblox.com address. - Roblox’s response: none so far, according to Open3DLab.
Watch: our 30-second video summary (in Vietnamese).
What Happened in the Roblox AI Scraping Incident?
Open3DLab runs three free 3D community sites with a small team: SFMLab (since 2014), SmutBase (since 2016, adult content only), and Open3DLab itself (since 2020). Creators upload characters, props, and maps, pick a license for each project, and anyone can download them for free.
Here is the Roblox AI scraping timeline Open3DLab describes:
| Date (2026) | What happened |
|---|---|
| Oct 15–16 | Crawler reads the main page of 18,926 projects in 26,429 requests over 34 hours, then downloads 1,027 files |
| Oct 17–18 | Activity pauses over the weekend |
| Oct 18 | Open3DLab spots it, blocks it, and emails the address in the user agent |
| Oct 19 | Crawler returns; every request is blocked |
| Oct 20 | Final attempt from 8:00 until about 15:00, tapering after lunch |
| Oct 26 | Open3DLab publishes its report; still no reply from Roblox |
Two details of the Roblox AI scraping stand out. All requests came from a single Amazon Web Services address. And the activity followed office hours in Central European time, with a pause around lunch. The only thing that capped the download was the speed limit on the sites’ free download servers. Even so, the scrape saturated those servers for a couple of days and caused problems for ordinary users.
Did the Crawler Really Belong to Roblox?
This is the most important caveat in the Roblox AI scraping story, and Open3DLab states it openly. The Roblox link rests on the user agent, which is a self-description. Anyone can type any email address into a user agent, and the crawler ran from Amazon’s cloud, not Roblox’s own network.
Open3DLab says it saw nothing that suggests impersonation, and that the label fits Roblox’s published research closely. It still lists confirmation as the first question it wants Roblox to answer. The site also says it does not blame the individual researcher, and it removed the person’s name from the report. In its view, a data-collection run like this is a company decision.
On Hacker News, where the post reached 475 points and 266 comments, one commenter argued the timing looked like a single person running a script on a laptop. Another replied that a personal project would hardly use a work email address during working hours.
How the Roblox AI Scraping Links to Roblox Cube
Open3DLab marks this part of the Roblox AI scraping report as its own speculation, based on public information:
- Roblox Cube. In March 2025, Roblox introduced Cube, which it calls its core generative AI system for 3D and 4D. The paper describes the goal as a “foundation model for 3D intelligence.” It lists character rigging as a future capability and generating characters and avatars as future work.
- Cube’s training data. Its first shape model was trained on roughly 1.5 million 3D assets from licensed and public datasets such as Objaverse, plus opted-in free assets from Roblox’s own Creator Store.
- 4D generation. In February 2026, Roblox began rolling out “4D generation” built on Cube.
- The researcher. The email in the user agent belongs to one of the 40-plus authors of the Cube paper, whose public profile says they are building a 3D foundation model at Roblox.
- “trellis2.” The name may refer to TRELLIS.2, an open 3D generation model Microsoft released in late 2025. Open3DLab notes that TRELLIS.2’s published data toolkit has no web crawler, so this crawler was written for the job. It might also just be a project codename.
Why would these sites matter for Roblox AI scraping? Open3DLab says they hold what public 3D datasets lack: complete, textured, often rigged characters in the formats artists use, each with a title, description, and preview images. Its conclusion is that the crawler was most likely collecting training or evaluation data for Roblox’s 3D model work. It also says it cannot prove this, or that any file reached a model.
Why the Adult Content Matters
By Roblox’s own figures, more than a third of its daily users were under 13 in 2025, and its rules ban sexual content. Yet the crawler took 224 projects from SmutBase and 121 adult-marked projects from SFMLab, at the same rate as everything else.
Open3DLab says it does not think Roblox set out to collect adult content, and it has no evidence any of it reached a product. Its point is that nobody seemed to check. If adult data reaches a model that generates assets, the only safeguard left is whatever filter is built afterwards.
There is a second problem: deletion. When a creator removes a project, or the site takes one down after a copyright notice, it disappears from the site but stays in any copied dataset. One downloaded project has already been disabled on SFMLab.
The Other Side of the Argument
Not everyone on Hacker News agreed with how the Roblox AI scraping was framed. Several commenters suggested adult data could just as well be used to train a filter that detects adult uploads, which a platform for children would need. Others called the focus on adult content sensational. Some argued that anything downloadable will be downloaded, while others replied that ignoring licenses and robots.txt is exactly what makes it wrong. Critics also pointed out that many models on the sites are fan ports of game characters. Open3DLab admits this in its post and says it complies with takedown notices.
Is Roblox the Only AI Crawler?
No. Open3DLab says that in the month before it found the Roblox-labeled crawler, Meta’s crawler, Meta-ExternalAgent, made about 380,000 requests and loaded 211,000 project pages. Open3DLab has blocked it too. The difference, it says, is that the Roblox-labeled crawler is the first one it saw actually download the files users uploaded.
What the Roblox AI Scraping Case Means for Creators
The Roblox AI scraping lesson for anyone who shares work online is uncomfortable: robots.txt and a license are requests, not locks. Here, a crawler fetched the robots.txt file and then ignored it.
Practical points raised in the report and the discussion:
- Rate limits matter. Speed caps on free servers were the only thing that limited the Roblox AI scraping download.
- Watch your logs. Open3DLab found the crawler because of its unusual request volume.
- Blocking cloud ranges is a trade-off. Some commenters block traffic from AWS and other clouds, but others warned this can also block real users on corporate VPNs and legitimate search engine crawlers.
- Keep evidence. Open3DLab kept a full export of the relevant logs for outside auditing.
For more on how AI systems handle other people’s data, read our look at AI chatbot privacy trackers and our report on rogue AI agent security failures.
FAQ
Did Roblox confirm it ran the crawler?
No. Open3DLab says it emailed Roblox twice about the Roblox AI scraping and got no reply. The Roblox link comes from the crawler’s user agent, which Open3DLab treats at face value.
How much data was downloaded?
1,027 files from 712 projects, 145.5 GB in total, on October 15–16, 2026, according to Open3DLab.
Was the data used to train Roblox Cube?
Nobody knows. Open3DLab thinks training or evaluation data for Roblox’s 3D foundation model work is the most likely explanation, but says it cannot prove it.
Does robots.txt stop AI crawlers?
Only if the crawler chooses to obey it. In the Roblox AI scraping case the crawler retrieved robots.txt and ignored it.
What does Open3DLab want from Roblox?
A public response, and for Roblox to respect the community’s norms: credit people, respect the license they chose, and ask before using work for something it wasn’t offered for.
Sources
- Open3DLab, “Roblox downloaded 145 GB of models from our sites”
- Hacker News discussion Last updated: October 28, 2026