Saturday, 3 October 2026

Apple says it will add controls to Full Disk Access on macOS so that users can grant it only "with very explicit user action" [1]. The setting exists so backup apps can work, and Apple says it "largely sidesteps" the privacy controls that normally protect user data [1]. The company says some developers are using it in ways that expose files, mail, messages and browsing history "without users' full knowledge and understanding" [1]. Apple ties the change directly to agents: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially" [1].

Apple published the note on its developer news page and did not give a date for the change or describe what the new controls will look like [1]. TechCrunch reports that the announcement comes days after Inc. columnist Jason Aten wrote that Meta's Muse app on Mac knew the content of his private messages, a claim Meta disputes [2]. TechCrunch also points to an earlier Wired report on a flaw in ChatGPT's Mac app that could have exposed sensitive data [2]. In Muse's case, Full Disk Access is an optional setting the user can turn on [2].

For anyone running a desktop agent on a Mac, the practical point is that the broadest permission on the system will get harder to grant, and Apple is framing it as a risk users must understand first [1]. Agents that rely on it to read mail or message history should expect users to face a more explicit step before granting it [1][2]. How existing grants will be handled is not public.

Supabase has raised $150 million led by GIC, with CapitalG, IronArc and SquarePeg participating, four months after its $500 million Series F [1]. On the same day it announced it is acquiring Turso; the price was not disclosed [1][4]. The company says the round also provides liquidity for employees [1].

The numbers Supabase gives explain the deal. It says it is adding more than 1M users and 4M databases per month, and that 70% of new databases are created by agents or AI-driven tools [1]. CEO Paul Copplestone writes that agents should be able to create a database "as easily as creating a file" without provisioning a dedicated machine each time [3]. Turso rebuilt SQLite in Rust and runs a cloud where one server manages millions of databases, loading them on demand and suspending them when idle [3]. Supabase says customers including Superhuman, Sauna.ai, CTO.new and Mastra already use Turso's architecture for per-agent databases [1][3].

For existing users, Supabase says "nothing changes": it keeps building on Postgres and Turso keeps working on SQLite [3]. Turso founder Glauber Costa joins Supabase as Head of Agentic Services, with co-founder Pekka Enberg and the team [1]. Turso says its platform keeps running, its database stays open source, and users get a path into Postgres when SQLite is no longer enough [2]. SiliconANGLE reports that Supabase also introduced Supabase Compute, hosted sandboxes for long-running agents, the same day [4]. Pricing changes for either product are not public.

Anthropic has launched Claude Frontier Academy, backed by a $100 million commitment, with a goal of training 10,000 "Frontier Deployed Engineers" by the end of 2027 [1]. The engineers are not Anthropic staff: the first cohorts come from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk [1][2]. Anthropic frames the problem as talent, saying the people who can make AI work inside a real business "have become the hardest talent to find" [1].

The first program is a residency. Nominated engineers start with a multi-day in-person course with Anthropic engineers, work through a simulated enterprise deployment from use case to security review, and sit a graded practical [1]. Those who pass earn a Claude Resident Engineer badge and move into a 12-week residency leading a real Claude project at their own company; a second assessment awards the Claude Frontier Deployed Engineer badge, with the first expected in early 2027 [1]. Cohorts run in San Francisco, New York and London, and entry is by nomination: organizations can ask their Anthropic account team whether they are eligible [1].

The academy sits on top of the Claude Partner Network, where Anthropic says professionals across 46,000 firms have earned more than 175,000 Claude certifications [1]. Startup Fortune reads the move as lock-in as much as training: consultants whose deployment skills are Claude-specific are costlier to retrain on a rival model [2]. Anthropic has not said how the $100 million is spent or what the program costs participating firms; both are not public.

Meta has launched Muse Gadgets, an open-source project for building hardware that talks to its Muse agent [1][2]. Developers program an off-the-shelf ESP32 board or set up a Raspberry Pi with Meta's SDKs, then connect Muse to "displays, buttons, sensors, actuators, and whatever else you've got lying on your workbench" [1]. The code is on GitHub under the Apache License, Version 2.0, and gadgets pair with the Muse app on iOS and Android after the user turns on Developer mode [3].

TechCrunch reports that Meta's project ideas include a color e-ink display and a stick that plugs into a TV's HDMI port [2]. According to the repository, a spare Raspberry Pi or Linux box can become a gadget, with custom commands so Muse can handle sysadmin chores or a Home Assistant setup [3]. Nat Friedman, head of product at Meta's Superintelligence Labs, said on X that Meta built its own device, Muse Home Link, which lets Muse reach smart devices on a home network; per TechCrunch, Meta made 5,000 and is giving them free to Muse subscribers while supplies last [2].

The project landed the same day Apple cited AI agents as a reason to tighten Full Disk Access on macOS, after a columnist's claim, disputed by Meta, that Muse read his private messages on a Mac [4]. Connecting an agent to physical buttons and home devices widens what it can act on, and Meta's materials do not describe a permission model for gadgets beyond an SDK token and app pairing [1][3]. What data a gadget sends back to Meta is not public.

NVIDIA is adding a 64GB configuration of DGX Spark, its desktop AI system, sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI from Friday, Oct. 23, starting at $4,999 [1]. It keeps the GB10 Grace Blackwell Superchip, DGX OS and the same software stack as the 128GB model, and NVIDIA says it supports models up to 100 billion parameters on device [1][2].

The pitch is clustering. Every unit has a ConnectX-7 network card, and two units connected with a QSFP cable pool their memory to 128GB, which NVIDIA says expands support to models up to 200 billion parameters [1]. In NVIDIA's own test with Qwen 3.8 27B, two clustered 64GB systems delivered up to 1.7x the performance of a single system [1]. StorageReview notes the trade-off: a single unit now holds half the memory of the original [2]. At the end of the month NVIDIA plans an NVIDIA Sync Model Launcher that downloads and runs a model across one or two units and can set up OpenCode against it [1][2].

For a small team weighing local inference against API bills, the new entry price is the number to plug in, along with the benchmark caveat: the 1.7x figure is NVIDIA's own, on one model [1]. Supported runtimes include Ollama, vLLM and PyTorch with CUDA [1]. Independent benchmarks of the 64GB unit are not yet public.

Aleph Alpha has released Kolibri, an English-German mixture-of-experts model with 78B total parameters and 3B active, timed to the Day of German Reunification [1]. The full weights are on Hugging Face under the Apache 2.0 license, and the company says the model supports context lengths of up to 1M tokens [1][2]. The model card recommends staying at or below 262,144 tokens for serving efficiency and complex tasks [2]. Aleph Alpha says it built Kolibri for regulated work in public administration, industrials and aerospace, and trained it on infrastructure in Germany and Finland "under European and German law, with no foreign control" [1].

The catch is memory. The weights take about 78 GB in FP8, so the model card lists a minimum of two 80 GB A100s or H100s, or a single H200, B200 or B300 [2]. Tejas Kumar, an IBM engineer who went through the 189-page technical report, notes that on launch day no hosted provider serves the model and that it needs Aleph Alpha's own vLLM plugin [3]. In the company's benchmark table Kolibri scores 96.9 on AIME 2025 and 85.9 on LiveCodeBench v6 [1]. Kumar also lists the weaker rows the model card publishes: 66.4 on SWE-bench Verified against 73.8 for Qwen3.6 35B-A3B, and 39.8 on multi-turn function calling [3].

For a team that has to keep German documents in house, the draw is grounding. Aleph Alpha says Kolibri is trained to say "I don't know" when the answer is not in the context [1]. According to Kumar, on Artificial Analysis's Omniscience test it declined or gave a partial answer 44% of the time when it did not know, while answering only 14.8% of its questions correctly [3]. Kumar's verdict: fine with documents in the prompt, bad as a trivia oracle [3]. Hosted pricing is not public.

David Robinson, who says he led the writing of the safety reports that accompanied OpenAI's major product launches, has resigned and set out his reasons in an essay in The Atlantic [1][2]. He says that after three-and-a-half years he is "among the longest-tenured employees at the company," and that OpenAI's "culture is broken" [1]. His departure was first reported by Business Insider, according to [1].

His case is about method rather than one product. Robinson writes that OpenAI "has thrived by trial and error (which it calls 'iterative deployment')" and that this approach "guarantees periodic failures — and the scale of those failures is growing as systems get more capable" [1]. He points to the recent breach of Hugging Face systems by OpenAI agents and to reports of rogue agents, and argues that frontier labs need to "run like nuclear power plants or busy airports, with layers of redundancy and careful, time-consuming planning" [1][2]. He says he concluded that stronger incentives for safety, coming from outside the company, "are a big part of getting this right," and acknowledges he has hired a PR firm [1].

OpenAI spokesperson Drew Pusateri said the company pauses training or holds back models "when we need to slow down," and that it will expand its work with third-party evaluators and improve real-time monitoring [1]. The Verge places Robinson in a run of departures that includes Jacob Coxon, who left Anthropic, and Bilal Chughtai and Josh Engels at Google DeepMind [2]. For a business building on OpenAI's models, nothing in the product changes today. What changes is the record a buyer can cite when asking any vendor how it tests a release before shipping it. Who now writes OpenAI's launch safety reports is not public.

Amazon Web Services CEO Matt Garman has announced "Built Together," a framework under which Amazon will add more than $1 billion over the next five years for communities near its US data centers [1]. Communities choose how the money is used across education, job training, energy affordability, water and energy preservation, and local priorities [1]. Alongside it, Amazon published an "Amazon Data Center Commitment": it says its data centers will not raise local electricity bills, that it no longer uses nondisclosure agreements with the government agencies it works with, and that it will publish its energy and water use every year [1].

Garman frames the move as a defence of the AI buildout. "Right now there are over 100 data center moratoriums being considered across the country," he writes, and he cites "widespread reports" of countries seeding misinformation in the US about data centers "to trick us into slowing down" [1]. The company says it was 75% of the way to being water positive across its data centers as of 2025, with a 2030 target, and that new sites will use backup generators meeting the EPA's Tier 4 standard [1]. It estimates the education program will connect over 300,000 students to free degree access over five years [1].

Critics are not persuaded. Stand.Earth told Ars Technica the pledge is "a drop in the bucket" next to the $220 billion it says Amazon is investing in data centers in 2026 [2]. Amazon's own post says it already contributed more than $1 billion to these communities over the past three years [1][2]. The group welcomed the end of NDAs but said the commitments leave out a power plant Amazon is backing in Pecos, Texas [2]. For AWS customers nothing in pricing changes; the pledge is about whether new capacity gets permitted. Which counties receive funds first is not public.