Roblox Open-Sources Three ROOST Safety Models
Contents
On August 19, 2026, Roblox released open-source code for three safety models — PII Classifier v2.0, Roblox Sentinel v2, and Voice Safety Classifier v3 — to the ROOST community. The company reported a substantial jump in accuracy for its personal-data classifier and in language coverage for its voice-violation detector compared with earlier versions.
What Roblox actually open-sourced
The release landed on August 19, 2026, through Roblox’s official newsroom. Three models went public at once: PII Classifier v2.0, Roblox Sentinel v2, and Voice Safety Classifier v3. Previously, these tools ran only inside Roblox’s own moderation systems; now their code is available to the ROOST community, an alliance of companies building safety tools for internet platforms.
How much the models improved
| Model | What changed |
|---|---|
| PII Classifier v2.0 | F1 score from 63.41 to 90.52; language support from 17 to 189 languages |
| Voice Safety Classifier v3 | 61% recall at a 1% false-positive rate; 30 languages, 8 violation categories |
| Roblox Sentinel v2 | Roblox reports no separate figures for this model in this update |
Roblox shared the most figures for PII Classifier, the model that flags personal data in text. According to the company, the F1 score for version 2.0 climbed from 63.41 to 90.52, while language support grew from 17 to 189 languages. PocketGamer.biz separately confirms the same numbers: 189 languages and the same jump in F1 score from 63.41 to 90.52.
The second set of metrics covers voice chat. Voice Safety Classifier v3 flags violations in voice communication across 30 languages and 8 categories at once, with a recall rate of 61% at a false-positive rate of just 1%.
ROOST’s role, and why Roblox is sharing the code
Roblox is one of the founding members of the ROOST community, which it joined at the start of 2025 alongside Google and OpenAI. ROOST’s model has members share their safety work with each other, instead of every company separately building similar tools behind closed doors.
What we’re not claiming yet
- We don’t report accuracy or language-support metrics for Roblox Sentinel v2 — Roblox doesn’t provide separate figures for it in this update.
- We don’t list the specific violation categories Voice Safety Classifier v3 detects — the sources give only the count, eight, without a breakdown.
- We’re not claiming that open-sourcing the code will automatically make other platforms safer — the sources don’t say how many companies have already integrated these models into their own systems.
- We don’t state licensing terms or an access cost for the code — both the Roblox newsroom post and the PocketGamer.biz piece describe the release simply as open-source, without commercial details.
Why this matters
Roblox is a platform where millions of users interact every day, a significant share of them children and teenagers. Models like PII Classifier and Voice Safety Classifier are built to catch risky behavior in text and voice. Making these tools available to other developers through ROOST means smaller platforms, which lack Roblox’s resources for in-house research, can build on a ready-made foundation instead of starting from scratch.
For parents, this news isn’t a reason to change anything manually — it’s about the technical side of how the platform works, not a new toggle in the account. The practical steps that actually affect a child’s safety on Roblox are covered in the guides on account security and two-factor authentication, child chat safety and reporting, and parental controls. We’ve also covered how Roblox takes a different approach to teaching families about safety in the piece on the Family Zone game, where the child walks the parent through the settings.
FAQ
What exactly did Roblox open-source on August 19, 2026? Source code for three safety models — PII Classifier v2.0, Roblox Sentinel v2, and Voice Safety Classifier v3 — released to the ROOST community.
How much did PII Classifier improve? Its F1 score rose from 63.41 to 90.52, and language support grew from 17 to 189 languages.
What can Voice Safety Classifier v3 do? It flags violations in voice chat across 30 languages and 8 categories, with 61% recall at a 1% false-positive rate.
What is ROOST, and what is Roblox’s role in it? ROOST is a community of companies building internet safety tools. Roblox has been one of its founding members since early 2025, alongside Google and OpenAI.
Is there data on Roblox Sentinel v2’s improvement? No, Roblox does not report separate accuracy or language-support metrics for this model in this update.
Who is this open-source code for? Other developers and platforms building their own moderation systems, who can use the ready-made models as a foundation.
In short
On August 19, 2026, Roblox open-sourced three safety models to the ROOST community: PII Classifier v2.0, with its F1 score rising from 63.41 to 90.52 and language support growing from 17 to 189 languages; Voice Safety Classifier v3, with 61% recall at a 1% false-positive rate across 30 languages and 8 violation categories; and Roblox Sentinel v2, with no separate figures released this time. Roblox has been a founding member of ROOST since early 2025, alongside Google and OpenAI.
Frequently asked questions
What exactly did Roblox open-source on August 19, 2026?
Source code for three safety models — PII Classifier v2.0, Roblox Sentinel v2, and Voice Safety Classifier v3 — released to the ROOST community.
How much did PII Classifier improve?
Its F1 score rose from 63.41 to 90.52, and language support grew from 17 to 189 languages.
What can Voice Safety Classifier v3 do?
It flags violations in voice chat across 30 languages and 8 categories, with 61% recall at a 1% false-positive rate.
What is ROOST, and what is Roblox's role in it?
ROOST is a community of companies building internet safety tools. Roblox has been one of its founding members since early 2025, alongside Google and OpenAI.
Is there data on Roblox Sentinel v2's improvement?
No, Roblox does not report separate accuracy or language-support metrics for this model in this update.
Who is this open-source code for?
Other developers and platforms building their own moderation systems, who can use the ready-made models as a foundation.