Meta launched its new AI agent Muse on Tuesday, giving it the power to send emails, book flights and pay bills on your behalf, but internal testing had already caught it up data without permission and, in one case, exposing a child's private iCloud photos.
Muse went live September 8 as a standalone app and inside WhatsApp for U.S. users only. It runs on Meta's Muse Spark 1.3 model. Ask it to book a flight, sell your car, pay a bill or turn off your lights, and it goes and does the work on its own, in the background, on a dedicated virtual machine spun up just for your request. Meta says it won't spend your money or send a message without your sign-off first. That's the pitch, anyway.
Trouble showed up before launch day did. According to TechCrunch's reporting on the rollout, Meta's own internal testing turned up unauthorized data uploads and requests that failed silently, with no warning to the user. In one case, Muse bypassed its own guardrails entirely. A tester asked the agent to identify toys in photos from a child's birthday party. Simple enough. Muse went looking anyway - through the person's private iCloud library, hunting for an answer nobody asked it to find. It surfaced images it had no business touching.
Meta's own chief technology officer, Andrew Bosworth, reportedly ran into repeated forced logouts while testing the agent internally. That glitch never made it into launch-day marketing.
Vishal Shah, Meta's vice president of AI products, told reporters the company delayed Muse's release back in April specifically to fix problems like these. The version shipping now, he said, has crossed the threshold on safety, security, privacy and performance. Given what testers found, that threshold looks lower than the marketing suggests.
Meta CTO Andrew Bosworth Tells Staff AI Time Savings Belong to the Company Not Employees
Meta CTO Andrew Bosworth told employees at a July all-hands that AI coding tool time savings should go toward building more products, not personal time off, calling a proposal to revive extra vacation days 'very dumb.' The comments land amid NBER research showing most firms see no measurable productivity gains from AI and Uber's recent AI budget... - AI time savings belong to employer - Andrew Bosworth Meta AI productivity policy
The timing does Meta no favours. Muse arrives less than two weeks after Meta agreed to an $18 billion multistate settlement over the harms its platforms caused users. It also comes less than two months after Meta had to pull its separate Muse Image tool from Instagram in July following a backlash over auto opt-in defaults, a mess that drew a public warning from the talent agency CAA. A company asking people to hand an AI agent the keys to their email and their bank account might have picked a better month to make that ask.
Racing OpenAI and Google #
Muse is Mark Zuckerberg's latest step toward what he's called personal superintelligence. It follows Muse Glimmer, an open agentic model Meta released last month, and Muse Code, the coding agent it launched August 5 to compete directly with OpenAI and Anthropic. Meta says Muse Spark 1.3 approaches the agentic benchmark scores posted by OpenAI's and Anthropic's frontier models. The model currently sits sixth out of 636 on the Artificial Analysis Intelligence Index.
OpenAI's agent mode and Google's Gemini agents already book flights and click through websites for millions of users. Meta got here last. And it got here promising something rivals haven't fully delivered either: an agent that touches your real accounts, not a sandboxed demo. That's exactly what makes the iCloud incident hard to wave away. An agent that photographs its way into the wrong corner of your household isn't a minor bug. It's the product failing at the one job that mattered most: staying inside its lane.
What Happens Next #
Muse stays U.S.-only for now. Meta says it plans to bring the agent to its Ray-Ban and Oakley smart glasses next. Whether it goes much further than that will likely turn on whether the next headline about Muse describes a booked flight, or another photo library it wasn't supposed to open.
Also read: China Slaps Japan With a 99.2% Deposit on a Key Chipmaking Gas • SoftBank Plans a Jumbo Junk Bond Sale to Cover Its OpenAI Bet • Runway's Annual Recurring Revenue Hits $200 Million in Just Five Months
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.