The day's tech, sifted: Oct 3, 2026

Sat, Oct 3

What matters today: Apple is tightening Full Disk Access on macOS so apps can only get it through very explicit user action, two weeks after a columnist said Meta's Muse agent surfaced one of his private message threads. Meta, meanwhile, open sourced code for DIY Muse gadgets and Amazon pledged $1B for data center communities. Yesterday OpenAI's agents were the security story; today the fix lands at the operating system layer, while OpenAI's own safety bench keeps thinning: Safety Systems author David Robinson went public in The Atlantic after resigning.

AI / LLMs

  • Meta open sources Muse gadgets: firmware for ESP32 boards and a Linux SDK (Raspberry Pi 5 included) that lets you attach its Muse agent to your own displays, buttons and sensors. It turns the agent into a hobbyist platform, days after the Messages controversy below.
  • Meta lets go of employees it hired from AI safety startup Virtue AI four months after they joined in June, citing clashing work styles.
  • Leaked Slack messages show OpenAI staff pushed back in June on Greg Brockman's ties to the Leading the Future group, contributing to him reneging on a $25M donation. Separately, Safety Systems leader David Robinson left OpenAI last week. Update: he has since gone public (next entry).
  • Update on Robinson: the man who wrote the safety reports that shipped with every major OpenAI model release resigned this week and, in an Atlantic editorial, says the industry's culture is fundamentally broken, a deeper problem than a few new rules can fix. The Verge cautions against dismissing such warnings because their authors helped build the thing.
  • OpenAI launches GPT-6.1 Sol at one-fifth of Astra's price, per TestingCatalog. Endor Labs says on Codex it matches Astra on security while running a third faster, a vendor-adjacent claim worth waiting on independent numbers for.
  • Aleph Alpha releases Kolibri, an open-weight model for German and English pitched as sovereign European AI. It drew 312 points on Hacker News.
  • OpenAI's DevDay push, per TechCrunch's read, turns ChatGPT into a place to discover, launch and use software, which could disrupt the app store model.
  • DeepSeek and Huawei release open-source Ascend programming tools, including compute and communication libraries, to make Huawei chips a workable alternative to Nvidia's software stack.

Devtools & Infra

  • Zig 0.17.0 ships, a long release-notes read that also led the front page tonight.
  • Supabase raises $150M led by Singapore's GIC and agrees to acquire Turso, a database built for AI agents, so the Postgres vendor is buying into agent-native storage.
  • Prime Intellect launches Prime Inference, already serving 600B tokens a day, using prefill/decode disaggregation (splitting the two phases of generation across machines) and NVFP4 KV compression.
  • Nvidia adds a 64 GB DGX Spark at $4,999, $1,000 above the 128 GB model's launch price: the memory shortage now shapes local-AI hardware.

Security & Privacy

  • Apple limits Full Disk Access (the macOS permission that lets an app read nearly everything) so users must take very explicit action to grant it, citing a "substantial" rise in agent risk (developer note). The trigger was columnist Jason Aten's claim that Muse referenced his Messages thread; Meta says Messages access is opt-in and needs both full-disk access and a Messages connector, so the dispute is over how easily users grant both.
  • A US judge ruled that an officer's license plate search in Flock's network was "indiscriminate mass surveillance" violating the Fourth Amendment, saying the officer should have got a warrant. It is a precedent that could limit warrantless use of plate-reader databases.
  • A URL-encoding bypass in Cisco SD-WAN (software that manages company wide-area networks) grants unauthenticated admin access, and US federal agencies were given 24 hours to respond, which signals active exploitation risk.
  • Vercel's Guillermo Rauch says its sandbox bounty program confirmed a KVM zero-day, a flaw in Linux's virtualization layer, a reminder that the isolation underneath agent sandboxes is itself a target.
  • The US arrested tech CEO Greg Lui, accused of using false paperwork and South Asian freight forwarders to divert more than $300M of servers with Nvidia A100 and H100 chips to China. He is charged; the claims are allegations.

Startups & Industry

  • Amazon pledges more than $1B over five years to communities near data centers, promising no higher power bills and "water positive" by 2030, and says it has stopped using NDAs with county officials. The pledge drew fresh criticism, and Amazon itself acknowledges backlash is driving local moratoriums.
  • Lyft settles for $272.5M with California's attorney general and three city attorneys over claims it misclassified drivers as contractors from 2016 to 2020. Uber's case in the same 2020 suit continues.
  • Netflix is ending some multiyear deals with prestige directors such as Fincher and Levy, which The Verge reads as a strategic shift as Paramount and Warner Bros. Discovery near a megamerger.

Research

  • An AI system has beaten top human players at Stratego, a game of hidden information, and did it cheaply, per Ars Technica. Imperfect-information games are closer to real decisions than chess or Go.

Elsewhere

Hacker News

Language and tooling lead. Zig 0.17.0 drew 227 points and 152 comments, while Greg Kroah-Hartman's talk on security in the LLM age hit 201. Developers are weighing the agent stack: a month of coding with GLM 5.3 Flash, the Redis creator's local LLM runner ds4, and essays on agentic coding's four horsemen and SaaS becoming a harness around a model. Aleph Alpha's Kolibri topped the evening at 312 points, and Apple's Full Disk Access change and Muse Gadgets also ran, covered above.

On privacy plumbing, Cloudflare's OHTTP gateway and Kagi's Orion update for Linux and Windows found audiences. Outside code, Linux on the M4 and a 12-year image sequence of planets orbiting a star did well, and an essay on social reality in China drew 165 comments.

Threads