Roblox open-sources Sentinel to protect kids in chat

Roblox has open-sourced Sentinel, its AI safety tool designed to spot grooming and child endangerment by analyzing chat patterns.

Rather than relying on basic keyword filters, it was built to catch the subtle, coded, and indirect language that predators use to evade human moderators.

What it does

Sentinel acts as an automated safety watchdog for live conversation feeds.

Inside Roblox, the system actively processes over six billion chat messages every single day in real time.

So far in 2025, it has helped detect 1,200 potential child exploitation cases that were reported to the National Center for Missing and Exploited Children (NCMEC).

Its biggest strength is proactive intervention. Roblox reports that in the first half of 2025, 35% of detected cases were flagged by Sentinel before any user or parent ever submitted an abuse report.

How it works

Sentinel analyzes context over time rather than judging isolated sentences.

The engine operates on one-minute snapshots of chat activity. By evaluating conversational dynamics across that window, it flags suspicious exchanges and tracks evolving patterns for deeper review.

The system was trained directly on past violations of Roblox’s child-endangerment policies, allowing it to benchmark live dialogue against real-world predatory tactics.

Roblox runs Sentinel alongside its broader moderation stack and recent protections like Trusted Connections, continually adapting the setup as bad actors invent new evasion methods.

Why it matters

For developers building online games, social spaces, or apps where kids hang out, Sentinel offers a battle-tested safety blueprint that is rare to see open-sourced.

Traditional blocklists fail against modern grooming, yet engineering an AI architecture capable of monitoring billions of live messages is normally out of reach for smaller teams. Having access to Roblox's production-grade pattern detection lowers that hurdle instantly.

That said, deploying it won't be plug-and-play. Exploitative language shifts constantly, meaning teams will still need ongoing updates and human reviewers in the loop. Even so, proving that proactive AI can catch over a third of dangerous cases before a victim even reports them resets the baseline for chat safety across the industry.