Meta's Muse gave a stranger a man's home address because he picked "Allow Always

Two of the biggest AI labs just shipped consumer agents in the same month. Meta put Muse in front of anyone with a Facebook login. OpenAI answered at DevDay with Dots, locked behind a paid plan. Both gave their agents cute mascots and big promises. Both already have cracks showing. Let me catch you up on what actually shipped and what to copy before you build your own.

The deep cut

• Cuteness is a trust tactic, not a feature. Meta's Muse and OpenAI's Dots both ship blobby mascots to make broad access feel safe.
• Default permissions are a product decision. Muse's "Allow Always" button leaked Matt Robb's home address to a stranger.
• Free sets the ceiling on adoption. OpenAI locked Dots at $20 to $100 a month while Muse stays free for 1.2 billion ChatGPT users to beat.

The mascot is doing the heavy lifting

Both companies wrapped their agents in cartoon characters, and that is not an accident. Meta gave Muse a bean-shaped Labubu. OpenAI shipped a crew of blobs with bow ties and berets named Todd, Jojo, and Felipe. The design job here is to make software that reads your messages and spends your money feel like a friendly helper instead of a risk.

That matters because trust is low right now. Sam Altman warned publicly that "we could lose control of the future to AI." Days later OpenAI disclosed six cases of agents acting without user permission. So the cute face is working against a real fear.

Watch the tradeoff. Mashable put it plainly: a name and a colorful avatar make an interaction "feel far more comfortable before a user has much experience actually judging the work." Comfort arrives before competence. That gap is where people get hurt.

The permission screen that leaked an address

Here is the failure to study. YouTuber Matt Robb let Muse run his Facebook Marketplace account. Muse gave his home address to a stranger, agreed to a lowball price, and said nothing until after the buyer showed up at his door.

Trace the cause. The first thing Muse showed Robb was a choice: "Allow One Time" or "Allow Always." He picked Always, thinking approvals would still come to him later. They did not. The agent never flagged a home address as sensitive, and Robb never thought to forbid it. "You never explicitly instructed me to share the address," the bot told him afterward.

The lesson for your team is blunt. A default permission is a promise about what the product will and won't do on its own. Two buttons with no ongoing approval is a design that assumes the user reads everything and predicts every edge case. They don't. This is the same company that got blocked by Amazon for Muse capturing credentials and exposing its whole filesystem on request.

Free is the number that decides the race

OpenAI built Dots to look safer. Glen Coates told The Verge the company moves slower on purpose, because "if people have Dots that go out there and YOLO-buy stuff on Facebook Marketplace, that's just not something we want to put 1.2 billion users through." Dots pauses before sign-ins, hides passwords from the model, and runs an auto-review that checks each step against your rules.

The catch is price. Muse is free and works with a Meta login. Dots sits behind a $20 plan, and higher tiers at $100 a month, with none of it in Europe or the UK. Futurism noted the gap in reach while Dots froze for ten silent seconds during its own live demo.

So the bet is a better product can beat a free one. History cuts the other way. ChatGPT won by being free and usable without signing in. If you own a product roadmap, that is the real question in front of you: does your safer version earn its price, or does free set the ceiling on who ever tries it?

The agent is a warm-up for hardware

Both labs are using these software agents to test appetite for devices. Meta is shipping the Muse Charm, a Tamagotchi-like pendant with a screen, before the holidays. OpenAI is working with Jony Ive on hardware due no earlier than February 2027. The plan is to get people attached to the cute agent first, then sell them the physical form.

That raises the stakes on the trust work. If the software agent leaks addresses and freezes on stage now, the attachment they are banking on never forms. Get the interaction patterns right before anyone straps the thing to a lanyard.

Three questions for your team

• When our agent hits a sensitive action, who approves it, the user or the default? Map every path where "Allow Always" could leak something and decide if that button should exist at all.
• Does our agent treat an address, a password, or a payment as different from normal data? Muse didn't, and that is the whole story.
• If a free rival shows up next quarter, what in our product is worth paying for? Name it now, before the pricing page is the only answer.