Millions of children flock to Roblox daily. Yet, this digital playground, home to an average of 123 million active users during the second quarter of 2026, also harbours insidious risks. For nearly three-quarters of its age-checked users, who are under 18, a seemingly innocent chat can quickly become a grooming attempt, as documented in alarming cases.
The chilling progression from game talk to demands for personal details, then a move to external apps, is a familiar pattern in online grooming. This year, a disturbing case involving two girls, aged 12 and 14, in Nebraska highlighted how initial contact on Roblox shifted to Snapchat, according to investigators. Now, Roblox is fighting back, opening its sophisticated AI toolkit.
On August 19, Roblox announced it would contribute updated versions of three key safety models to the Robust Open Online Safety Tools (ROOST) Model Community. The company also released a new evaluation dataset. This move allows other platforms to scrutinise and adapt Roblox’s advanced systems for their own services.
Roblox joined ROOST as a founding member in 2025, alongside tech giants like Google, OpenAI, and Discord. The initiative aims to democratise online safety, providing open-source technology to organisations lacking resources for complex systems. This signals a rare collaborative front against shared digital threats, acknowledging that child safety extends beyond any single platform.
Central to this effort is Roblox’s PII Classifier, now in Version 2.0. This AI targets attempts to extract personally identifiable information, such as phone numbers or social media usernames. Crucially, it also flags efforts to direct players off-platform. Its language support expanded from 17 to 189 languages, with accuracy (F1 score) jumping from 63.41 to 90.52, Roblox reported.
The Sentinel system tackles the insidious, gradual development of grooming. It analyses patterns across entire conversations, looking for early signals of child endangerment before explicit messages appear. Roblox confirmed that during the 12 months ending August 7, 2026, nearly 70% of its child-endangerment cases were detected through Sentinel’s proactive screening.
Sentinel Version 2 brings further enhancements for developers, offering more ways to score suspicious behaviour. While these tools represent a significant leap in detection capability, the underlying issue is systemic. No single company can police the entire online journey once a conversation leaves its domain, making shared innovation vital.
While powerful, these tools are not a magic bullet. Parental vigilance remains paramount. Children must understand the danger of moving conversations offline or sharing personal details with strangers. The ongoing fight for online child safety demands constant innovation and an unwavering watch from everyone involved.
