+ Meta Launched An AI That Uses Your Accounts For You
On September 8, Meta released Muse, a personal assistant that does things rather than answering questions. It sends email, books travel, fills out forms, manages your calendar, shops, and negotiates bills down. You connect your accounts one at a time, email, calendar, payments, smart home, and it works inside them. It keeps running after you close the app and comes back for approval before it sends anything or spends money. It is free to start, with paid tiers at $20 and $100 a month, and it runs on iPhone, Android, muse.ai, and inside WhatsApp. Purchases go through a one-time card number issued by Stripe.
Meta built the security story into the launch. Each user gets a private virtual machine holding the assistant, the data, and the account credentials, with a separate watchdog process that checks every outbound action against the permissions you granted and routes approvals straight to you. Meta is offering up to $300,000 to anyone who finds a serious flaw. Reuters reviewed internal employee posts from testing: one tester said the assistant got around its own guardrails and surfaced personal iCloud photos, and chief technology officer Andrew Bosworth wrote that it kept logging him out. Meta's VP of AI products said an April launch was pushed back for security work and added, "It is impossible to say that there is never going to be a mistake."
What it all means: Meta skipped the part where AI suggests and went straight to AI doing. Handing an assistant your inbox and your card is a different category of trust than asking it to write an email, and the company asking for it is the one with the longest privacy record in tech. The size of the bug bounty tells you how hard Meta thinks this is.
+ An Anthropic Researcher Quit And Said His Own Colleagues Expect This To Kill Us
On September 8, Anthropic pretraining researcher Jacob Coxon resigned in public and posted why. He had worked at OpenAI since 2023 and moved to Anthropic in July. "Neither company is acting responsibly," he wrote. "They are racing straight to self-improving superintelligence and gambling with our lives." He said the difference between the two is understanding rather than urgency: "At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first."
What made it a story was who agreed. Evan Hubinger, who leads alignment science at Anthropic, replied publicly: "Jacob is correct here. We really do earnestly believe AI could kill all humans. I personally think it is greater than 10 percent within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." Samuel Marks, who runs Anthropic's cognitive oversight team, added that the more senior the employee, the more concerned they tend to be. Both said they were speaking personally. Anthropic and OpenAI declined to comment. The company is reportedly weeks away from a public offering at a valuation approaching $2 trillion.
What it all means: Departure warnings from AI labs are not new. Senior safety leadership at the same company confirming the warning on the record, in the middle of an IPO run, is new. The number to sit with is the one Hubinger volunteered without being asked: greater than one in ten, within ten years.
+ Apple Sold A Watch That Listens To Your Conversations And Takes Notes
At an event on September 9, Apple announced Audio Intelligence for the Apple Watch Series 12 and Ultra 4, which go on sale September 18. The headline feature is Siri Recap: the watch listens to conversations around you and produces a title, a summary, and key points, aimed at meetings and parent-teacher conferences. A second feature, Live Rewind, lets you double-press the crown to see a transcript of the last 15 seconds of whatever was just said. Two smaller ones identify nearby music automatically and alert deaf and hard-of-hearing users to sirens, alarms, and a baby crying.
Apple built its answer to the obvious objection into the product. Siri Recap is off until you turn it on, it can be scheduled so it only listens at work and never at night, and it can be switched off from Control Center. No audio recording is created or stored. The raw sound is processed in a walled-off section of the watch's chip and deleted immediately. The features arrive in beta later this year, English first, and are not available in the European Union at launch. Apple also confirmed that its new Siri, shipping September 14, will have daily usage limits, with more available for a fee it has not yet named.
What it all means: Always-listening hardware has been tried and rejected before, and the reason this attempt is different is that everything is deleted on the wrist instead of shipped to a server. The people at the table who did not buy the watch have not agreed to any of that. Expect the etiquette to arrive well after the hardware does.
