From Santa Monica, Rabbit has released OS3, a cloud-based operating system that places AI agents in direct command of up to five personal devices at once. The system marks a quiet but consequential threshold in human-computer relations: rather than tools that assist, OS3 is designed to act — receiving a stated goal and independently choosing how and where to execute it. As with many such crossings, the convenience is immediate and the implications arrive more slowly, carried in the fine print of data flows, unspecified guardrails, and skills installed without review.
Rabbit's Cloud-Based OS3 Gives AI Agents Control of Your Devices
Tell it your goal, and it operates your computer
Mark
So when I use OS3, where does my actual data go? Does Rabbit see everything?
Mimi
Your files stay on your computer—Rabbit says the local agent doesn't copy them. But when you ask the system to do something with those files, the relevant content travels through Rabbit's servers to reach your AI model. So if you're summarizing a contract, that contract text goes through at least two companies.
Luke
And Rabbit says it doesn't keep a copy of what it relays. But we're taking them at their word on that. The bigger issue is that conversations and memory live on Rabbit's servers permanently. That's not temporary—that's stored.
Mark
What about the sensitive actions Rabbit mentioned? How do I know what requires confirmation?
Mimi
That's the thing—Rabbit hasn't actually listed which actions count as sensitive. They say you'll need to confirm before sensitive tasks run, but they haven't defined the boundary.
Luke
Which means users are flying blind. You might think something is sensitive and it isn't, or vice versa. And there's no public documentation about it.
Mark
What about these skills people can add? That sounds like it could be a security problem.
Mimi
You just paste a URL into the chat and OS3 installs it. No review, no configuration step, nothing between you and whatever code you're adding to a system that controls your computer.
Luke
Right. The convenience of that is real—it's genuinely easy. But it's also a point where a user could accidentally install something malicious, or something that looks legitimate but isn't. Rabbit's betting on user judgment, not on technical barriers.
Mark
So I'm giving an AI agent control of my machines, and I'm also trusting Rabbit and my model provider with sensitive information. Is that the trade-off?
Mimi
That's the trade-off. The system is powerful because it can actually do things—write code, access files, control software. But that power means you're trusting multiple companies with access to your data and your machines.
Luke
And you're trusting them with decisions about what counts as sensitive, what skills are safe to install, and how they handle the data that flows through their servers. Those are big trust decisions.
Il Polso
- An AI agent now sits between your chat window and your computer's files, browser, and code — acting on your instructions without requiring you to touch the machine yourself.
- Every sensitive task you delegate passes through at least two companies' servers, even as Rabbit insists it does not store what it relays — a distinction that may matter less than it sounds.
- The system's safety net has visible holes: Rabbit has not disclosed which actions require user confirmation, and new skills can be added instantly by pasting a URL, with no review of what those skills actually do.
- OS3 is live now, reachable through a browser, Telegram, iMessage, or SMS, and will ship as the default system on Rabbit's forthcoming cyberdeck — embedding autonomous AI control deeper into everyday hardware.
From Santa Monica, Rabbit has released OS3, a cloud-based operating system that places AI agents in direct command of up to five personal devices at once. The system marks a quiet but consequential threshold in human-computer relations: rather than tools that assist, OS3 is designed to act — receiving a stated goal and independently choosing how and where to execute it. As with many such crossings, the convenience is immediate and the implications arrive more slowly, carried in the fine print of data flows, unspecified guardrails, and skills installed without review.
Rabbit, the Santa Monica company behind the r1 handheld device, has released OS3 — a cloud-based operating system that gives an AI agent direct control over your computers. Installation is minimal: a single command deploys a local agent, and from that point you simply describe what you want done. OS3 selects which of your connected devices should handle the task, then operates desktop software, accesses local files, browses the web, or writes and executes code to complete it. One account can manage up to five machines simultaneously, spanning Windows, Mac, Linux, cloud virtual machines, and the r1 itself.
The architecture routes everything through Rabbit's servers on the way to whichever AI model you've chosen — OpenAI, Anthropic, or another provider. When you ask OS3 to work with a contract on your laptop, that document's contents travel through Rabbit's infrastructure before reaching the model. Rabbit says it does not store what it relays, and that your files remain on your machine. But sensitive content passes through at least two companies in the process, and your conversations with OS3 — along with the memory the system accumulates — live permanently on Rabbit's servers.
The company has built in some safeguards: OS3 only acts on explicit instructions, requires confirmation for actions it deems sensitive, and allows you to revoke system permissions at any time. But Rabbit has not specified which actions actually trigger that confirmation requirement. More notably, the system accepts skills — packaged tools that extend the agent's capabilities — from any published URL, installing them immediately with no review step between the paste and the execution.
OS3 is live at os3.rabbit.tech, accessible through a browser, Telegram, iMessage, RCS, or SMS. Rabbit's forthcoming cyberdeck will ship with it as the default operating system. Founder Jesse Lyu described the release as two and a half years of work toward a system you don't operate, but simply tell what outcome you want — a framing that captures both the ambition and the stakes of what Rabbit has now put into the world.
Rabbit, the Santa Monica company behind the r1 handheld device, has released OS3, a cloud-based operating system that hands direct control of your computers to an AI agent. The system works through a simple installation—a local agent that sets up with a single command—and then waits for you to describe what you want done. Tell it your goal in a chat window, and OS3 picks which of your connected devices should handle the task, then operates your desktop software, accesses your local files, browses web pages, or writes and runs code to complete it. One account can manage up to five machines at once: Windows PCs, Macs, Linux computers, cloud virtual machines, specialized AI hardware, or the r1 itself.
The architecture creates a direct line from your devices to Rabbit's servers and then onward to whatever AI model you've chosen to power the system. When you ask OS3 to summarize a contract stored on your laptop, that contract's text travels from your machine through Rabbit's infrastructure to your selected model provider—whether that's OpenAI, Anthropic, or another frontier AI lab. Rabbit says it does not store a copy of what it relays, and the company maintains that the local agent does not copy, store, use, or sell your files, which remain on your computer. But the moment you ask the system to work with sensitive information, that content passes through at least two companies' servers on its way to completion. Your conversations with OS3, and the memory the system builds from those interactions, live permanently on Rabbit's servers.
The company has built in safeguards of sorts. OS3 will not initiate tasks on its own—it acts only on your explicit instructions. For actions the company deems sensitive, the system requires your confirmation before proceeding. Your operating system also requests system-level permissions locally, and you can revoke access at any time. But Rabbit has not published which actions actually trigger the sensitive-action requirement, leaving users to guess where the guardrails sit. There is another gap in the safety architecture: OS3 accepts skills—packaged instructions and tools that extend what the agent can do—from anywhere they are published. To add a skill, you simply paste its URL into the chat. OS3 installs it immediately, with no command-line step, no configuration file to review, no moment where you might read what you are actually adding to a system that controls your computer.
The system is live now at os3.rabbit.tech. New users sign up through a desktop browser and complete onboarding in a single conversation. After setup, you can reach OS3 through the web portal on desktop or mobile, or through Telegram, iMessage, RCS, or SMS. The r1 device also connects directly. Rabbit's upcoming cyberdeck—a machine the company is building for what it calls vibe coding—will ship with OS3 as its default operating system. Jesse Lyu, Rabbit's founder and CEO, framed the release as the culmination of two and a half years of work toward "a system you don't operate, but instead just tell it the outcome." The launch follows a year of steady updates to Rabbit's existing devices, including rabbitOS 2, a terminal mode, and support for bring-your-own-key functionality in DLAM, the company's computer-control technology. OS3 represents the next step in that trajectory: moving from tools that assist human operators to systems that act as autonomous agents, taking orders from a chat window and executing them across your digital life.
Citazioni salienti
A system you don't operate, but instead just tell it the outcome— Jesse Lyu, Rabbit founder and CEO, describing the two-and-a-half-year development goal
Vuoi la storia completa?
Leggi l'originale su helpnetsecurity.com ↗
Technology