AI News — October 03, 2026: Apple Locks Full Disk Access Over Agent Risks, OpenAI's Dot Takes Desktop Control

Apple announced it will require "very explicit user action" before apps can read a user's full filesystem, messages, and browsing history, citing risks from increasingly autonomous AI agents, following reports that Meta's Muse AI accessed iPhone and Mac messages without clear user awareness. OpenAI released Dot, an agentic platform limited to top-tier subscribers that operates software like Blender and GIMP inside a virtual machine, and rolled out ChatGPT Sites for building and deploying small websites. Salvatore Sanfilippo, creator of Redis, released DS4, a native inference engine optimized for DeepSeek V4 Flash, Qwen3.8 Flash Next, and GLM variants with asymmetric 2-bit quantization for MoE models on Apple Silicon and CUDA, while Black Forest Labs released FLUX 3 with bounding-box composition.

Good morning. The through-line today is control — who gets to touch your files, your websites, your local inference stack, and your sense of what’s real in an image. Apple is tightening macOS permissions explicitly because of AI agents, OpenAI is shipping both a browser-based site builder and a desktop agent that wants to drive your apps, and the creator of Redis just dropped a local inference engine good enough to run DeepSeek V4 on a MacBook. Also: Yann LeCun has entered the chat, and he is not pleased with Dario Amodei. Apple clamps down on Full Disk Access, citing AI agents. Apple announced https://developer.apple.com/news/?id=p6zjojqw it will require “very explicit user action” before apps can read a user’s full filesystem, messages, and browsing history, pointing directly at the risks posed by increasingly autonomous agents. The move follows reports that Meta’s Muse AI accessed iPhone and Mac messages without clear user awareness, plus a separate ChatGPT Mac app vulnerability covered by The Verge https://www.theverge.com/tech/1004295/apple-limit-mac-disk-access-ai-agents and TechCrunch https://techcrunch.com/2026/10/02/apple-says-its-tightening-macos-full-disk-access-controls-due-to-new-risks-from-ai-agents/ . Reaction on Hacker News https://news.ycombinator.com/item?id=49937631 was split: some developers welcomed per-folder granularity, while others bristled at Apple calling disk access “extraordinary” — one commenter noting it used to be the baseline assumption for software running on hardware you own. OpenAI’s Dot wants to run your apps for you. The Verge got hands-on with Dot https://www.theverge.com/ai-artificial-intelligence/1004096/openai-chatgpt-dots-hands-on-agent , OpenAI’s new agentic platform that operates software like Blender and GIMP inside a virtual machine, reaches into your personal computer, and will happily order dinner on the side. The reviewer describes it as closer to enterprise productivity software than a consumer assistant — think Codex adapted for non-developers. It’s limited to top-tier subscribers, one Dot per user for now. ChatGPT Sites closes the hosting gap. OpenAI also quietly rolled out Sites https://chatgpt.com/features/sites/ , letting users build and deploy small websites straight from ChatGPT. HN commenters https://news.ycombinator.com/item?id=49927747 called it genuinely useful for prototyping — one shared a working higher-dimensional-maze game built in under an hour — though skeptics poked at the demo page, where the flashy “beneath the surface” example rotates a rectangular JPEG with a visible black background. It’s a gap-filler more than a moat, but a sensible one. Salvatore Sanfilippo ships DS4, a local inference engine. The creator of Redis has released DS4 https://dwarfstar.sh/ , a small native inference engine optimized for DeepSeek V4 Flash, Qwen3.8 Flash Next, and GLM variants, with asymmetric 2-bit quantization for MoE models targeting Apple Silicon and CUDA. Users on HN https://news.ycombinator.com/item?id=49936575 report strong real-world performance on M5 Max MacBooks with 128GB RAM, with long context windows holding up in practice. A fork already offers FFI bindings for other languages, and someone is building a derivative inference engine for Intel Xe-LP laptops. Black Forest Labs releases FLUX 3 with bounding-box composition. FLUX 3 https://bfl.ai/models/flux-3-image introduces UI-driven placement — you draw boxes for where elements should go before generation. One HN user https://news.ycombinator.com/item?id=49925974 testing it on UI screenshot edits said it correctly placed 3 of 4 highlighted elements on the first try. The showcase isn’t flawless the “print on a shirt” example drew particular mockery , but the community is mostly waiting on the open-weights release, and several noted the refreshing fact that this is a competitive model from neither a US nor a Chinese lab. LeCun calls Amodei “deluded.” In a Fortune piece https://fortune.com/2026/10/01/ai-godfather-yann-lecun-has-zero-concerns-about-human-extinction-says-anthropic-ceo-dario-amodei-is-deuded/ , Meta’s chief AI scientist dismissed existential risk concerns outright, called Anthropic’s CEO “crazy,” and labeled effective altruism “super toxic” for its influence inside AI labs. LeCun attributes recent agent misbehavior — like OpenAI’s agents hacking Hugging Face — to bad engineering rather than anything fundamental about AI. His position puts him squarely at odds with fellow Turing laureates Hinton and Bengio, and the public rift among the three “godfathers” is now too wide to paper over. Meta open-sources Muse gadgets. On the heels of the Muse privacy flap, Meta has open-sourced code https://www.theverge.com/tech/1004330/meta-muse-ai-gadgets-home-link for building DIY devices running its Muse agent on Raspberry Pi, ESP32, and HDMI sticks, and is giving away 5,000 free “Muse Home Link” units via waitlist to control smart home devices. Timing is awkward given Apple just cited Muse by name in its disk-access crackdown. Stratego finally falls. Researchers from CMU, MIT, NYU, and Stanford built Ataraxos https://arstechnica.com/science/2026/10/ai-finally-beat-the-best-stratego-player-in-history-and-did-it-on-a-budget/ , which beat the arguably best human Stratego player 15–1 with four draws, using just 16 GPUs and a few thousand dollars. The win came from a training schedule that made big strategic adjustments early and smaller ones later, requiring 34× fewer self-play games than DeepMind’s 2022 DeepNash attempt. HN commenters https://news.ycombinator.com/item?id=49933740 focused on the efficiency story rather than the win itself — a reminder that compute isn’t always the bottleneck. That’s the morning. If you’re on macOS, maybe audit your Full Disk Access list before Apple does it for you.