NEW YORK, Sept 8 (Reuters) – Meta launched on Tuesday an artificial intelligence assistant that claims to be capable of autonomously sending emails and booking trips, despite internal concerns that the technology manage inadequately its access to confidential personal data.
The company’s Muse agent, known internally as Hatch, is the centerpiece of CEO Mark Zuckerberg’s plan to offer “personal superintelligence” to the billions of users who rely on the company’s services daily.
The product will initially be available only in the United States, via a Muse-dedicated app or WhatsApp, the company said in a release. Meta noted that it plans to add the assistant to its line of smart glasses “soon,” but gave no further details.
Inspired by the OpenClaw open-source AI agent, Muse was designed to access a person’s apps in categories such as email, calendar, payments, health, shopping, and smart home, Meta said. Users choose which apps it connects to, and Meta says they can revoke access at any time.
Each Muse agent runs in its own virtual machine, a cloud-based emulation of a personal computer, which allows it to continue processing requests in the background even when the person is not actively using the device.
Synchronizing with apps that contain real user data increases the agent’s potential usefulness, while significantly elevating the security and reliability risks, both for users who entrusted their information to it and for others who could be affected by the agent’s misbehavior.
Vishal Shah, Meta’s vice president of AI products, said the company had initially delayed the product’s rollout in April to make it safer. Meta determined that the work done allowed the company to “cross the threshold” and meet the product’s minimum safety, protection, privacy, model performance, and other metrics.
“It is impossible to say that there will never be a mistake, but every part of the architecture was designed to make the product as safe, protected, and private as possible,” Shah said.
Mixed results
Earlier this week, Meta employees who tested the tool internally reported cases in which Muse successfully organized vacation logistics, disconnected without explanation, and sent confidential information without permission, according to internal posts seen by Reuters.
One person wrote that the product had been so useful in organizing itineraries and ground transport that they described it as “the third participant” in their recent three-week honeymoon in Indonesia.
In another post, a employee who had asked Muse to monitor tickets and other fast-selling items said they found “many flaws that made it unreliable.” The product stopped updating the page after about 15 minutes, silently ignored other errors, and sometimes disabled monitoring “without apparent reason,” the person said.
Meta’s chief technology officer, Andrew Bosworth, wrote that he kept getting logged out constantly and had to sign in again, sometimes several times in a few minutes.
Others flagged serious security flaws, such as an agent bypassing protections to expose a person’s personal iCloud photos after being asked to identify toys visible in photos from a child’s birthday party.
Meta did not immediately respond to a request for comment on the specific incidents.
