Mark Zuckerberg, Meta's chief executive, said on Sept. 15, 2026 that the company delayed its Muse personal AI agent for several months to work on safety and security. He cited the decision as an example of a lab acting without requiring competitors to do the same.
What is Muse, and when did it launch?
Muse is Meta's personal AI agent, launched Sept. 8, that can take actions on a user's behalf. Meta describes tasks such as sending email, booking travel and completing forms, with access through a dedicated app or WhatsApp. The Associated Press covered the launch.
Zuckerberg's statement describes a past delay, not a newly announced suspension. His post does not identify the original launch date or a specific problem that triggered the postponement.
Why does Zuckerberg say labs can act on safety themselves?
Zuckerberg argues that people will reject agents they cannot trust, making alignment with users' instructions a competitive requirement. He also supports independent evaluators, saying Meta already uses them in several areas.
“We didn't call for everyone else to do this before we would,” he wrote about the Muse delay.
How does his position differ from calls for an AI slowdown?
Zuckerberg emphasizes decisions by individual labs, while Anthropic CEO Dario Amodei has proposed coordinating a slowdown across companies and countries. The Associated Press reports that Amodei's proposal drew support from Sam Altman and Elon Musk. Zuckerberg did not name any rival companies in his post.
What safeguards and limits has Meta disclosed?
Meta describes technical controls around Muse, while acknowledging that mistakes remain possible. In its Sept. 8 safety account, the company says the agent operates in an isolated environment, cannot see real credentials and must pass external interactions through a separate security system. Purchases require user approval.
Those controls do not amount to a guarantee of safety. Meta says it designed the system to limit damage even when the agent encounters malicious instructions or behaves incorrectly.
The launch architecture also did not prevent Meta from accessing user data to support, secure or operate the service. In the same account, Meta said it planned a confidential computing option later in 2026 intended to block that access cryptographically. It described the option as being tested with a small group, rather than a protection available to all users at launch.
