Back

Meta and OpenAI take different routes to keeping personal AI agents safe and private

Meta's diagram of Muse's design: the agent in a sealed runtime cell on a person's own virtual computer, beside a separate Sentinel that grants permissions and secure storage for login details
Image: Meta

Meta's Muse and OpenAI's dots, which OpenAI released Sept. 29, are both personal AI agents, but they guard people's data in different ways. Meta walls Muse off on its own secure computer, where a separate gatekeeper agent must approve anything it sends out, while OpenAI gives each dot its own cloud computer and checks risky actions automatically. How either design holds up once people connect their real work apps is not yet known.

Two personal AI agents, AI that can carry out tasks on a person's behalf, now take clearly different approaches to privacy. OpenAI released dots on Sept. 29. Meta's launch materials and safety research for its Muse agent describe another design: a sealed-off computer, a separate gatekeeper agent called Sentinel that alone approves outside actions, and passwords the agent never sees. For anyone deciding whether to let an agent near their accounts, the two companies are offering different kinds of protection.

Meta says Muse runs on Muse Secure VM, a dedicated, secure virtual computer with its own browser that holds both the agent and a person's data. Each person controls how much access Muse gets. Sentinel, a separate agent on the same machine, is the only one that can approve Muse's actions in connected apps or anything it sends to the internet: Muse proposes actions, and Sentinel decides which can go ahead, asking the person for permission when needed. Meta says Muse also checks with the person before sensitive actions such as sending an email or making a purchase. Meta says Muse cannot see people's passwords or payment methods, and login details go into secure storage so the agent can use them without seeing them.

Meta's AI research team says the design assumes the agent may be under attack. The agent runs in an isolated cell, and every interaction with the outside world passes through Sentinel, which the agent cannot override. Meta says it is opening a Muse bug bounty that pays up to $300,000 for valid reports of security flaws. That includes up to $130,000 for a successful prompt injection, an attempt to trick the agent with hidden instructions, that affects one user. Later in 2026, Meta says, it plans Muse Confidential VM, meant to encrypt the whole virtual computer, including a person's data and conversations, with a key only that person holds.

OpenAI's Sept. 29 post introducing dots describes different protections. Each dot works on its own cloud computer, and the user's own computer stays separate unless the user connects it. OpenAI says dots can use saved passwords for supported sites without exposing them to the model. For proactive research, which dots do without being asked, they reach connected apps only through read-only tools, so they cannot send messages, change what is in an app or control the browser or computer. Custom Rules let users decide when a dot acts, asks for approval or is blocked, and an automatic review checks actions that could affect accounts or share information. Sensitive tasks such as changing a password always stay with the user, and monitoring can pause or stop a dot over a safety concern.

Both designs are described so far only by the companies themselves. Nobody can yet say how Meta's Secure VM and planned Confidential VM hold up in everyday use, or how OpenAI's automatic review and read-only research behave once many people connect their real work apps.

More on OpenAI Meta