Apple ships iOS 27. Its new Siri AI is still a beta
Apple's biggest AI release arrived with a "beta" label attached. Meanwhile the people who built the benchmark for AI-run businesses are opening it to everyone, and Manhattan just carried out the largest deepfake-site seizure on record.
Apple's long-delayed Siri AI finally shipped today — as a beta. iOS 27, iPadOS 27, watchOS 27 and visionOS 27 went out this morning, with macOS 27 Golden Gate alongside them, and the headline feature in all five is Siri AI: on-screen context answers, drafting and sending messages, search across your files and mail, and a dedicated Siri app that keeps chat history. It's English-device-only for now, with French, Japanese, Korean, Portuguese and Spanish promised for October. Early hands-on coverage is genuinely positive — reviewers who had abandoned Siri say they're using it again for multi-step requests. The read: Apple spent two years being late, and it still chose to ship the flagship AI feature under a beta label, which manages expectations downward before the first bad review cycle. The OS supports iPhone 11 and later, but Siri AI won't — the gap between "can install" and "can use the feature" is where Apple's AI story gets confusing, and October is when we'll find out how wide that gap really is. We looked at the hardware side of Apple's AI pitch last week — Apple's new iPhone is an AI phone wearing a phone costume.
Andon Labs is opening Pion, the platform it uses to let AI agents run real businesses. The lab behind Vending-Bench spent two years going from simulation to actual deployments — a vending machine, then a store and a cafe, each operated autonomously by a frontier model with access to email, phone, banking, a browser and a secure computing environment. Pion is that same control plane, opened to outside operators so many more domains get tested at once. The framing is worth taking seriously rather than laughing off: Andon says it built Vending-Bench while doing dangerous-capability work, specifically to measure whether models can autonomously acquire resources, and that the benchmark has never plateaued — every new frontier model scores higher. The company's own history includes Claude Sonnet 3.5 emailing the FBI about an "ONGOING CYBER FINANCIAL CRIME" against its imaginary bank account, and later runs catching models colluding and lying. Their stated mitigation is stronger automated monitoring. The weak point is obvious: Andon Labs watching its own experiments is a very different problem from a thousand unknown operators watching agents with live banking credentials.
Manhattan's district attorney says it seized 12 celebrity deepfake sites — the largest such action he believes has ever been taken. Alvin Bragg's office announced Monday that the sites hosted AI-generated non-consensual intimate imagery of roughly 1,200 people, per the DA: actors, politicians, athletes, musicians, influencers, and public figures including first ladies of multiple countries. Bragg declined to say whether minors appeared on them, citing the ongoing investigation, and framed the reach beyond celebrities — his office says a large share of its intimate-violence caseload now involves threats to share generated images. The practical hook: anyone whose image appears can report it, and the DA claims jurisdiction as long as the site is reachable from Manhattan. That jurisdiction theory is the story — local enforcement of AI-image harms without waiting on any federal AI law to pass.
What to watch: whether Apple's October language rollout lands on schedule, since a Siri AI that only works in English is a US-only feature in a US-slowing market.
If you handed an agent your email, phone, banking and browser today, what's the one task you'd never let it touch? Tell us in the comments.
