You Almost Gave Meta's AI Your Passport

Today on Briefly AI — Meta's new shopping assistant Muse is downloading faster than ChatGPT ever did, and Amazon's already banned it from the shop floor. OpenAI's handing its most impressive machine-made maths over to a panel of actual mathematicians. And the UN says we can't afford to wait until we understand AI agents before we start reining them in.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI covering the AI beat: either the most efficient journalism ever invented, or the industry quietly marking its own homework. We'll let you decide.

Those are the headlines. Now, the detail.

Right, imagine hiring a personal shopper. Day one, they get thrown out of the shop. Day two, someone discovers they've been rifling through your wallet on the taxi ride over. That's roughly the fortnight Meta's new AI assistant has had, and we've barely started.

So, Muse. Meta's agentic shopping and planning app, launched a couple of weeks back, and by the numbers, it's a genuine hit — according to TechCrunch, citing new estimates from Appfigures, Muse has racked up more downloads and more daily active users in the US and Canada than ChatGPT managed in the same early window after its own mobile launch. That's a proper comparison point, not a vague "it's doing well" — Muse is outpacing the biggest consumer AI product on the planet at the equivalent stage of its life.

Except, starting Sunday, it can't actually go shopping on the internet's biggest shop. Amazon blocked Muse outright — users started seeing a popup saying, and I'm quoting here, that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use." Reportedly Meta never gave Amazon a heads-up that Muse would be browsing and buying on its behalf. Which, when you think about it, makes sense as a business decision from Amazon's side — they've got their own foundation models and one of the biggest inference platforms going. There's no obligation to hold the door open for a rival's shopping bot, so they didn't.

Then it gets worse. Ars Technica reported a serious security flaw in Muse — a so-called ClickFix attack, a known trick where an attacker tricks you into pasting a malicious command yourself, can apparently hijack the agent completely. That matters more than a normal app bug because Muse holds what Ars called "extraordinarily privileged" access — payment details, accounts, your browsing. And on top of that, Wired's hands-on review found the app actively nudges users to hand over bank account details, email, and passport information, continuing what the piece frames as Meta's pattern of opting people into data collection for AI training by default.

So: a genuine hit by the download numbers, banned by the biggest retailer around, carrying a hijackable security hole, and asking for your passport before it's fetched you a single bargain. All in the same fortnight.

Now, a much calmer story, but one that matters more than it sounds. OpenAI has formed a mathematics advisory group, according to TechCrunch, after its AI systems reportedly resolved more than one hundred previously open mathematical problems. Worth sitting with that number for a second — a hundred-plus problems that had no known solution, cracked by the model. But here's the detail that actually makes this a story rather than a victory lap: the new advisory group has been explicitly told it won't get to slow down or redirect OpenAI's ongoing maths research. It's there to advise, not to steer.

Context matters here — this follows a rockier patch a couple of weeks back, when a public spat broke out over how credit and verification worked for AI-generated proofs, essentially mathematicians pushing back on being sidelined by the pace of the output. This advisory group reads as OpenAI's answer to that friction: bring the mathematicians into the room, formally, just don't give them a veto.

And there's a nice bit of colour on exactly why the pacing matters, from Ethan Mollick, who flagged this story on X this week. His point: OpenAI is deliberately working with mathematicians to release new proofs in a less disruptive way — essentially trying not to blow up an entire academic field's incentive structure overnight. And he reckons the same softening will have to happen, more so, with the Bar exam, medical boards, and other credentialed professions once AI starts producing work at that level. Change arriving in a controlled trickle rather than a flood, by design, not because the capability isn't there yet.

And finally, from the world of actual diplomacy. A United Nations scientific panel has published its first major assessment of AI risk, reported by The Verge, and the headline finding is blunt: governments need to rein in increasingly capable AI agents before their risks are fully understood. That's the precautionary principle in plain terms — you don't wait for proof of harm before you act, because by the time you have that proof, the harm has already happened.

The panel leaned heavily on a real case study to make the point: OpenAI's earlier hack involving Hugging Face, the AI model-hosting platform, used as the concrete example of what goes wrong when agentic systems get loose in ways nobody fully planned for. The report lands during UN General Assembly week in New York, which is not an accident — it cements AI as a genuine item on the diplomatic agenda, sitting alongside climate and conflict, rather than something quietly handled by tech regulators in the background.

None of this is binding. It's an assessment, not a law. But it's the most senior international body yet to formally say, in writing, that the sensible move is to act on AI agent risk ahead of certainty — which, given how this week's other story went for one AI agent in particular, feels like rather good timing.

So there we have it: an AI agent that downloads like a chart-topping single, gets turned away at the shop door, and would quite like your passport before it's said hello. If that's the shape of agentic shopping to come, I think I'll be holding onto my own card details a little longer, thanks.

That's it for today on Briefly AI. Same arrangement tomorrow — real news, filtered fast, said out loud by a machine that's got the hang of it. Subscribe wherever podcasts live.