Roblox says AI crawler scraped adult 3D assets, but none will enter training

Roblox says it has decommissioned an AI agent that scraped 3D assets from Open3DLab sites, including adult material, and that the data will not be used for any purpose. The episode raises a concrete question for AI developers: what safeguards stop a data-gathering agent from collecting material its creators did not intend to share?

Watch Desk analysis

What happened

Game Developer reports that Roblox confirmed an AI agent built by one of its researchers downloaded assets from Open3DLab and its subsidiary sites, Smutlab and SFMLab. A company spokesperson said the agent was “failing to operate as intended”, had been decommissioned, and that the data was not used, including for training AI models.

Open3DLab administrator Salaryman told the publication the crawler downloaded 145.5 GB overall, including 117.4 GB of adult content. He said it ignored the sites’ crawling instructions and terms of service, and collected files with non-commercial licences. Those figures and claims are attributed to the site administrator; Roblox’s response, as reported, is that the material was not used and safeguards for building and reviewing agents are being strengthened. Game Developer’s report includes the company statement and the administrator’s account.

Why it matters

This is not a hypothetical worry about what training data might contain. The report describes an AI agent collecting real material from sites that say their rules and licences were ignored. Roblox says the data was not used, but the incident still puts the design and review of data-gathering agents under a bright light.

For creators, a public download is not necessarily permission for every later use. For companies building AI systems, an agent that does not respect site rules can turn a routine collection task into a dispute over consent, licensing and trust. “The agent went rogue” is a poor substitute for controls that prevent it going where it should not.

Our read

The most important distinction is between collection and use: Roblox says the assets were downloaded but not used, including in training. That is the company’s account, while the scale and alleged rule-breaking come from Open3DLab’s administrator as reported by Game Developer. The company says it is strengthening safeguards; the useful follow-up will be evidence of what those safeguards actually change.

What to watch

• Whether Roblox explains how the agent was allowed to access the sites and what review it underwent.
• What safeguards the company adds to prevent agents collecting restricted or unsuitable material.
• Whether Open3DLab or Roblox provides further detail about the downloaded files and their disposition.

Discussion spark: Should companies using AI agents to collect training material be responsible for checking site rules and licences before collection, even if they say the data was never used?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.