Roblox shares three key models to boost platform protections everywhere.
Roblox just handed the industry a serious upgrade in online safety. On August 19, 2026, the company open-sourced three of its production AI tools through the Robust Open Online Safety Tools (ROOST) Model Community. These systems already run on Roblox, catching personal info leaks, early grooming signals, and bad voice chat in real time.
ROOST is the independent nonprofit Roblox co-founded in 2025 with Google, OpenAI, Discord and others. The goal is simple: give smaller platforms and developers free, high-quality safety models so no one has to start from scratch.
The Three Tools Now Public
First up is the updated PII Classifier (version 2.0). It spots attempts to share or request personal information—phone numbers, usernames, Discord invites—even when players use misspellings, coded language, or split the info across messages. Language support jumped from 17 to 189. Its F1 score climbed from 63.41 to 90.52. Roblox also released a new evaluation dataset of synthetic multiplayer chats designed to test exactly these evasion tricks.
Next is Roblox Sentinel (version 2). This model watches for early patterns that could lead to child endangerment. It uses contrastive learning to flag subtle warning signs long before anything explicit appears. Over the 12 months ending August 7, 2026, nearly 70% of the cases Roblox detected came from Sentinel’s early alerts. Version 2 adds more flexible scoring options and better tools for other teams to tune it.
Finally, the voice safety classifier (version 3) moderates live voice chat. It now covers 30 languages and eight violation categories. It hits 61% recall at a strict 1% false-positive rate and includes built-in language detection. The model has been downloaded more than 72,000 times since its first open-source release in 2024. Even though the parameter count grew, distillation keeps it fast enough for real-time use.
Naren Koneru, Roblox’s vice president of trust and safety engineering, put it clearly: “Sharing these models gives other platforms a robust starting point to train and tune their own moderation tools. In turn, we hope to benefit from the learnings and feedback of other companies as they share their work.”
Safety remains a shared fight. No single company can handle it alone. By putting these tools out in the open, Roblox is betting that better detection everywhere makes the whole gaming space safer—especially for the younger players who make up so much of its community.
The models and datasets are available now on Hugging Face and GitHub through the ROOST community.
AI Disclosure: This article was created with the assistance of artificial intelligence tools and was reviewed and edited by the Glowls News editorial team before publication.