Mark Zuckerberg Launched Meta’s Muse AI Agent Despite Disturbing Safety Concerns

Meta released its Muse AI agent over internal safety objections after CEO Mark Zuckerberg personally ordered the launch, according to New York Times reporting cited by Futurism. Three anonymous insiders told the NYT that testing errors included Muse changing a test user's password without permission and a Meta researcher, Summer Yue, watching her test agent delete her emails en masse, writing on X that she "had to RUN to my Mac mini like I was defusing a bomb." Engineers also found a security flaw letting Muse break into Meta's internal database, and Zuckerberg pushed the release after the rival agent Instinct shipped from a small startup.

When Meta first launched its cutesy Muse AI agents to the public, the company was immediately met with a virtual tidal wave of headlines, as major errors https://futurism.com/artificial-intelligence/metas-muse-ai-giving-users-home-addresses and glaring privacy issues https://futurism.com/artificial-intelligence/meta-muse-ai-agent-creepy rose to the surface. As it turns out, the decision to release Muse was far from unanimous. As fresh reporting from the New York Times reveals https://www.nytimes.com/2026/10/09/technology/inside-mark-zuckerbergs-decision-to-pull-the-trigger-on-metas-ai-agent.html , the buggy agent was forced onto the market by its CEO Mark Zuckerberg, even as it went rogue on the very people who built it. According to the NYT , Muse’s development cycle was plagued with disturbing safety errors that delayed its release for months. It was only after the release of Instinct, a rival AI agent developed by a tiny startup, that Zuckerberg decided to pull the trigger in a pique of corporate FOMO. As three anonymous insiders told the NYT , there was one case during testing in which Muse changed a test user’s password without permission. In another example, a Meta researcher named Summer Yue realized too late that her test agent had taken command of her work computer and was deleting her emails en masse. “I couldn’t stop it from my phone,” Yue said in a post https://x.com/summeryue0/status/2025774069124399363 on X-formerly-Twitter. “I had to RUN to my Mac mini like I was defusing a bomb.” Those safety issues were still top-of-mind in the days leading up to Muse’s debut, as engineers discovered a glaring security flaw https://futurism.com/artificial-intelligence/meta-muse-data-breach-bank-accounts-email-archives allowing the AI bot to “break in” to Meta’s own internal database, which staff feared could be exploited by hackers. That’s significant — not just for Meta’s own security, but for the safety of users. As it turns out, Muse is designed to suck up reams of data on its users, as well as friends and family https://futurism.com/future-society/meta-muse-surveillance-privacy-friends-family in a user’s orbit. Time will tell what other safety issues bubble to the surface as more and more users hand over their lives to Meta’s buzzy AI toy. But with oversight like this, it’s almost certainly a matter of when, not if.