Roblox open-sources three safety models through ROOST

• The PII Classifier's F1 score also increased from 63.41 to 90.52.
• Roblox said Sentinel detected nearly 70% of its child-endangerment cases through early detection.
• The latest voice safety classifier supports 30 languages and eight violation categories.
• Roblox is also releasing a new benchmark dataset for PII classifiers.

Roblox plans to contribute three open-source safety models to the Robust Open Online Safety Tools (ROOST) Model Community.

The company joined ROOST as a founding member alongside Google, OpenAI and others in early 2025. The new safety models include updates to Roblox's open-source PII Classifier, Roblox Sentinel and its latest voice safety classifier.

Roblox said versions of the tools already operate on its platform to detect attempts to share personal information, identify early signs of child endangerment and moderate voice chat in real time.

The company also unveiled a new evaluation dataset designed to help other companies benchmark their own moderation systems.

Safety tools

The PII Classifier's latest version improves its ability to understand conversational context and detect attempts to bypass moderation through misspellings, coded language and obscure references.

Roblox Sentinel is designed to identify early warning signs of potential child endangerment. Roblox said nearly 70% of cases it detected during the 12 months ended August 7th, 2026 were identified through Sentinel's early detection.

Meanwhile, version 3 of its voice safety classifier supports 30 languages and eight violation categories, achieving 61% recall at a 1% false-positive rate.

Roblox said its new benchmark dataset also contains synthetic multiuser conversations designed to represent attempts to evade PII filters. Version 2.0 expands supported languages from 17 to 189 and raises its F1 score from 63.41 to 90.52.

Last week, we interviewed Magic Potion Games founder and CEO Stephen MacDonald about keeping Gen Alpha safe in online games.