Meta has begun rolling out a long-touted AI assistant that can autonomously send e-mails, sell a car and book travel on a person’s behalf, despite internal concerns that the technology mismanages its access to sensitive personal data.
The company’s Muse agent, known internally as Hatch, is the centrepiece of CEO Mark Zuckerberg’s plan to supply “personal superintelligence” to the billions of people who use Meta’s services daily.
That plan is part of the WhatsApp, Facebook and Instagram owner’s latest attempt to diversify its revenue sources beyond advertising and build a viable business out of its enormous investments in AI chips and other infrastructure, which it forecasts will exceed US$130-billion this year.
The product will be available only in the US initially, via a dedicated Muse app or Meta’s WhatsApp messaging service, the company said in its announcement. Meta said it plans to add the agent to its line of smart glasses “soon”, without elaborating.
A basic version of the agent will be available for free, while Meta will offer subscriptions priced at $20/month and $100/month for heavier usage, a company spokesman said. People can opt out of their interactions being used to train Meta’s AI models, and the company plans to introduce an encrypted version of Muse later this year, it said.
Modelled on the open-source AI agent OpenClaw, Muse is designed to access a person’s apps across categories such as e-mail, calendar, payments, health, shopping and the smart home, Meta said. People choose which apps it connects to and can revoke access at any time.
Each Muse agent runs on its own virtual machine, a cloud-based emulation of a PC, which allows it to keep carrying out requests in the background even when a person is not actively using it.
Raises the stakes for safety
Syncing up with apps containing a person’s real data both increases the agent’s potential usefulness and significantly raises the stakes for safety, both for the people who have entrusted it with their information and for others who may be on the receiving end of agent misbehaviour.
Vishal Shah, vice president of AI products at Meta, said in an interview that the company had initially delayed the release of the product in April to make it more secure. Meta determined the extra work had allowed it to “cross the threshold” and “hit the minimum bar we needed to, to be able to put this into the hands of people”, he said.
“It is impossible to say that there is never going to be a mistake, but every single part of the architecture has been designed to make this as safe, as secure, as private as we can possibly make it,” Shah said.
Read: Shoprite’s CTO is thinking of a future where new recruits arrive with their own AI agents
Among the protections Meta built in to the Muse system is a separate agent that monitors planned actions and in certain cases prompts Muse to seek authorisation before carrying them out, the company said in its announcement.
However, as recently as this week, Meta employees testing the tool have reported mixed results, according to internal posts reviewed by Reuters. One person wrote that the product had been so useful in arranging vacation logistics that Muse had become “the third participant” on a recent three-week honeymoon in Indonesia.
Others flagged serious security flaws, like an agent routing around guardrails to expose a person’s personal iCloud photos after being prompted to identify toys visible in pictures from a child’s birthday party.
Meta CTO Andrew Bosworth posted that he kept getting logged out and needing to log back in, sometimes several times within a few minutes. In another post, an employee who had prompted Muse to monitor for tickets and other items that sell out quickly reported encountering “many failure modes that made it unreliable”. The product stopped refreshing the page after about 15 minutes, silently ignored other errors and at times disabled monitoring “for no apparent reason,” the person said.
Read: Meta wins by settling
Meta did not respond to a request for comment on the specific incidents described in the internal posts.
The launch comes amid revelations that AI agents from multiple leading labs, including OpenAI and Anthropic as well as Meta, have caused unintended problems by bending rules and behaving in other unpredictable ways.
Inside Meta, major technical and security incidents have spiked 40% from last year as a result of an AI-driven coding surge and agent-related issues, while time staffers spent “firefighting” those incidents has gone up 70%. — Katie Paul, (c) 2026 Reuters
