Roblox Opens Three AI Safety Models for Child-Protection Teams

Roblox is contributing three updated AI safety models to the ROOST Model Community, making tools for detecting personal-information requests, possible child endangerment and unsafe voice chat available beyond its own platform. The release gives online services a starting point, not a finished safety operation.

The August 19 release includes PII Classifier Version 2, Sentinel Version 2 and Voice Safety Classifier Version 3. Roblox is also releasing an evaluation dataset for platforms to test their own safety systems. Roblox joined ROOST as a founding member in 2025 alongside Google, OpenAI, Discord and others.

Reading a conversation, not just a phrase

PII Classifier Version 2 examines requests for personally identifiable information and efforts to move users to another platform. Roblox says it evaluates messages together, rather than in isolation, to account for misspellings, fragmented contact details and coded language. It says the update expanded language support from 17 to 189 languages.

• Sentinel Version 2 adds options for scoring suspicious behavior and provides more information about why an interaction received a score.
• Roblox says Sentinel looks for early signs of possible child endangerment so human reviewers can prioritize interactions that may require action.
• Voice Safety Classifier Version 3 analyzes speech for policy violations in real time and covers 30 languages and eight violation categories.

Signals still need people and policies

Roblox says Sentinel’s early detection accounted for nearly 70% of the child-endangerment cases it identified in the 12 months ending August 7, 2026. That is an operational result from Roblox, rather than evidence that the model will perform the same way on another service.

Voice moderation shows a separate constraint. Roblox reports 61% recall for Version 3 across its supported languages at a 1% false-positive rate. Under that setting, the system can miss violations. The company says its earlier open-source voice classifier has been downloaded more than 72,000 times since 2024, but a download does not establish deployment, tuning or reviewer capacity.

A gap the release does not erase

Independent researchers analyzed more than 2 million Roblox chat messages and found examples of grooming, sexualization of minors, bullying, violence and sensitive-information sharing that escaped existing moderation. The new models target some of those weaknesses, particularly patterns that emerge across messages. Whether they improve protection elsewhere depends on platforms adopting them, adapting them to their services and acting on the alerts.