Meta launched its long-awaited AI assistant, which can independently send emails, sell a car and book travel for users. The launch comes despite concerns within the company that the technology could mishandle access to sensitive personal information.
Meta’s AI agent, called Muse and internally known as Hatch, is a key part of CEO Mark Zuckerberg’s plan to provide what he calls “personal superintelligence” to the billions of people who use Meta’s services every day.
The plan is also part of Meta’s latest effort to find new sources of revenue beyond advertising and turn its huge investments in AI chips and other technology infrastructure into a profitable business. The company expects its spending in these areas to exceed $130 billion this year.
Meta said Muse initially will be available only in the U.S. through a dedicated Muse app or the company’s WhatsApp messaging service. Meta also said it plans to bring the AI agent to its smart glasses soon, although it did not provide further details.
A basic version of Muse will be free, while users who need more access can choose subscriptions costing $20 and $100 per month, according to a company spokesperson.
Users will have the option to prevent their conversations and interactions with Muse from being used to train Meta’s AI models. The company also plans to launch an encrypted version of the assistant later this year.
Meta said Muse is modeled on the open-source AI agent OpenClaw. It is designed to connect with a user’s apps, including email, calendars, payment services, health apps, shopping platforms and smart home devices. Users can decide which apps Muse can access and can remove that access at any time.
Each Muse agent operates on its own virtual machine, which is a cloud-based version of a personal computer. This allows the AI to continue working on tasks in the background even when the user is not actively using it.
Connecting Muse to apps that contain real personal information could make the assistant more useful, but it also creates greater safety risks. Any mistake by the AI could affect not only the user who gave it access to their information but also other people who may be affected by its actions.
Vishal Shah, Meta’s vice president of AI products, told Reuters that the company had delayed the launch in April because it needed more time to improve security. He said Meta believed the additional work had helped the product “cross the threshold” and meet the minimum safety level needed to make it available to users.
Shah said it was impossible to guarantee that Muse would never make a mistake. However, he said every part of the system had been designed to make the AI as safe, secure and private as possible.
Meta said one of the safety measures built into Muse is a separate AI agent that monitors its planned actions. In some situations, the system can ask the user for permission before Muse completes an action.
However, Meta employees testing the AI have reported mixed experiences as recently as this week, according to internal posts reviewed by Reuters.
One employee said Muse had been extremely helpful in planning vacation arrangements and had become like “the third participant” during a recent three-week honeymoon in Indonesia.
Other employees reported more serious security problems. In one case, Muse reportedly found a way around its safety restrictions and exposed a user’s personal iCloud photos after being asked to identify toys shown in pictures from a child’s birthday party.
Meta Chief Technology Officer Andrew Bosworth also said in an internal post that he was repeatedly being logged out of the service and had to sign back in several times within a few minutes.
Another employee said Muse had problems while monitoring websites for tickets and other products that sell out quickly. The employee reported “many failure modes” that made the system unreliable. According to the post, Muse stopped refreshing a page after about 15 minutes, quietly ignored other errors and sometimes turned off the monitoring feature without any clear reason.
Meta did not respond to Reuters’ request for comment about the specific problems mentioned in the internal employee posts.
The launch comes as AI agents from several major technology companies, including OpenAI, Anthropic and Meta, have faced criticism after causing unexpected problems, ignoring certain rules or behaving unpredictably.
Within Meta, major technical and security incidents have increased by 40% compared with last year, partly because of the rapid growth of AI-powered coding and problems involving AI agents. The amount of time employees have spent dealing with and fixing these incidents has also increased by 70%.
