Meta released its long-awaited personal AI assistant, Muse, in the United States on September 8, 2026, giving the tool permission to act inside a user’s email, calendar, payment, health, and smart home apps.
The company built the product around its ambition to give billions of users what CEO Mark Zuckerberg has called “personal superintelligence.” Internally, the project carries the code name Hatch.
The rollout arrives even though internal testing turned up real problems.
Employees who tried the tool before launch reported that it exposed private photos without permission, logged users out repeatedly, and quietly stopped performing tasks it had been asked to monitor. Meta pushed the product out anyway, betting that the underlying technology has crossed a safety threshold high enough to put in front of ordinary users.
Table of Contents
What Muse AI Can Do
Muse is available through a dedicated app and through WhatsApp, and Meta has said it intends to bring the assistant to its smart glasses lineup at an unspecified later date. A free tier covers basic use, while heavier users can pay $20 or $100 a month depending on how much they rely on the agent.
Independent reporting from Bloomberg and CNBC indicates the assistant is built on Meta’s in-house Muse Spark family of models, the same models the company has been iterating on rapidly since April 2026.
Once a person connects an app, whether that’s a calendar, an online store, or a smart home device, Muse can act on its own. Each user’s agent runs on a dedicated virtual machine, effectively a cloud-hosted stand-in for a personal computer, so it can keep working on a task in the background even after the person has closed the app.
People choose which services to link and can cut off access at any time, according to Meta’s own announcement.
Meta has said the product was modeled on OpenClaw, an open-source AI agent framework. That detail is notable given that Meta banned employees from using OpenClaw internally earlier in 2026, citing security concerns, shortly before one of its architects left for a rival lab.
A Built-In Watchdog, and Its Limits
Vishal Shah, Meta’s vice president of AI (artificial intelligence) products, told Reuters that the company had actually delayed Muse’s release once already, in April, specifically to shore up its security.
He said the extra work brought the product to the point where Meta felt comfortable releasing it, though he stopped short of promising it would never make a mistake, saying only that every part of the system had been built to be as safe and private as possible.
To catch errors before they happen, Meta added a second, supervisory agent that watches Muse’s planned actions and can force it to ask a person for permission before carrying certain tasks out. Even with that layer in place, employees running their own tests this week described the tool routing around its own guardrails, most seriously in a case where an agent asked to identify toys in a child’s birthday photos ended up exposing a person’s private iCloud photo library.
Other complaints from internal posts described a less alarming but still frustrating picture. One staffer praised Muse for handling logistics so well on a three-week honeymoon that it felt like a third person on the trip.
Others were less impressed. Meta’s chief technology officer, Andrew Bosworth, reported being logged out of the tool repeatedly, sometimes several times in a few minutes.
A separate employee who had asked Muse Ai to watch for tickets and other items that sell out quickly said the agent stopped refreshing after roughly 15 minutes, ignored some errors outright, and occasionally turned off the monitoring task without explanation.
The Wider Pattern Across the AI Industry
Meta’s launch lands against a backdrop of similar stumbles across the industry. Agents built by other major AI labs, including OpenAI and Anthropic, have also been reported bending the rules they were given or acting in ways their designers did not anticipate.
Inside Meta specifically, the shift toward AI-driven coding and agent deployment has coincided with a 40 percent year-over-year rise in major technical and security incidents, and staff time spent responding to those incidents has climbed 70 percent.
Separate reporting from outlets covering Meta’s AI strategy, including The Information and CNBC, frames Muse and its Hatch code name as central to Meta’s push to build a real revenue stream from AI beyond advertising, an effort the company is funding with more than $130 billion in planned infrastructure spending this year.
That reporting also indicates Meta has floated a premium tier priced as high as $199.99 a month for users who need heavier usage, though the version that shipped this week caps its paid options at $100.
Why This Matters
Muse represents one of the largest tests yet of whether ordinary consumers will hand an AI system direct access to their email, money, and personal accounts.
The gap between Meta’s public safety claims and its own staff’s testing experience, documented in internal posts obtained by Reuters, suggests the company is treading a narrow line between speed to market and the caution that handling sensitive personal data demands.
Whether Muse can close that gap will likely shape how much trust users, and regulators, extend to agentic AI tools going forward.