How does Roblox use AI?
During internal testing, a hidden instruction in a GitHub issue convinced an AI coding agent to try uploading our credentials to a public repository. Traditional security tools saw nothing unusual. It happened in a sandboxed test environment we’d built for exactly that reason, and our security team later shared the full story at Black Hat 2026.
Roblox engineers use AI in almost every part of their work, and we test each new tool for how it could be misused before we roll it out. We want AI to take on the repetitive work so our teams can spend more of their time on design and judgment.
Our platform serves hundreds of millions of players, millions of creators building games and businesses, and billions of dollars across the ecosystem. In the U.S. alone, Roblox creators generated an estimated $752 million in GDP impact in 2025, a 69% increase over 2024. Many creators are building businesses and earning real income on Roblox, so security has to keep pace with everything we ship.
In the video above, you’ll hear directly from the people doing that work. Below is a closer look at the three ideas they keep coming back to.
Thoughtful adoption: Testing before we trust
Change can be unsettling, as frightening as it is exciting. As Principal Infrastructure Security Engineer Charles Zaffery shares in the video, “I think there was a period of time when I was really scared… this thing’s coming, it’s going to replace all our work. And now I don’t feel that at all.”
How we roll these tools out makes a big difference in whether our builders (the people who work here at Roblox) trust them. We treat AI adoption as an engineering and security problem to solve. We test each tool, build controls around it, and then trust it while we keep monitoring. The same security team that sandboxed our coding agents has built multiagent pipelines that research threats, find gaps in our detections, and write, challenge, and validate new detection rules, plus an autonomous threat-hunting system that runs against live security data.
Closer to everyday engineering, LLM-based pull request suggestions and agentic systems help us find and fix issues across our codebase. We also built an AI coding assistant trained on our own code, which nearly doubled the rate at which engineers accept its suggestions. That frees engineers to focus on high-impact design decisions instead of routine bug hunting. And because the best ideas rarely come from a roadmap, we put these tools in everyone’s hands. The fastest way to spark innovation is to make AI widely available and see where people take it.
Innovation at scale: Securing an ecosystem unlike any other
At Roblox, scale means 138 million daily active users, a message queue that handles more than 18 trillion messages a day at 99.995% availability, and infrastructure that spans owned data centers and cloud instances that spin up and disappear in seconds. It also means a live economy of millions of virtual items.
Not to mention that our business runs on code we didn’t write. Every day, Roblox executes scripts from millions of external creators. Some are brand new, some are seasoned, and occasionally a few are trying to get around our protections. So sandboxing, isolation, and runtime monitoring are built into the core of the platform.
As Senior Director of Engineering Dan Li explains in the video, our job is to build the framework that lets people code with confidence, with guardrails in place before surprises happen. AI helps us do that faster. Gungor Polatkan, Senior Director of Engineering for Discovery, notes that internal AI tools let engineers ask questions and understand existing products much, much faster. For Senior Software Engineer Sean Scarnecchia, that speed changes what teams are willing to attempt: AI lets them take on “the bigger bets [and] some of the harder parts of our roadmap with a little bit less hesitation.”
Consider secrets, the credentials and keys that keep our systems talking to each other safely: Across thousands of microservices, we built one central system that manages and rotates them automatically, so no engineer has to remember to refresh a credential. Our security teams treat engineers as their customers and work to make the secure option the easiest one to choose.
Respect the community: AI in service or human judgment
Protecting a platform this size starts with respecting the people on it, and that starts with an ethical choice: We don’t overconsume or overexpose data.
We have to move fast without cutting corners. Every AI integration is measured against our commitment to trust and reliability, and human judgment outranks any model’s recommendation.
Some people assume that more automation means less human connection. We’ve found that handing off mechanical work gives people more time for higher impact work and for each other. For Avni Yagnik, manager of Detection and Response, compassion is core to how we work. When AI carries the heavy lifting, people have more room for the nuanced, ethical calls that algorithms can’t make.
Come build with us
Polatkan sums it up best: “I don’t think I’ve ever worked in a place where this richness of AI is in literally every single aspect… This is what I was looking for.”
If you want to use AI to tackle some of the hardest problems in engineering, with the autonomy to push boundaries and a team that has your back when you take a big swing, we’d love to meet you.