Why it mattersIf you run AI agents on a Mac, giving them access to your files, messages and mail will take a more deliberate step, a sign that platforms now treat agents as a distinct security risk.
Apple says it is changing its macOS privacy settings to stop third-party app developers from misusing them to access message histories. — Ars Technica
Apple says it will add new controls around macOS’s Full Disk Access permission, warning that increasingly capable AI agents make broad access to users’ files, messages, mail, and browsing history riskier. — TechCrunch
Apple Inc. said it plans to introduce new privacy controls for Mac users, warning of the growing risks around granting broad data access to third-party software, including artificial intelligence agents. — Bloomberg
DeepMind researchers propose "Artificial Symbiotic Intelligence" as an alternative to the singularity, arguing general AI will emerge as a network of cooperating agents and humans rather than a single supermodel. They say the rules and institutions governing that cooperation will matter more than model size.
Harvard physicist Matthew Schwartz used the open-source BootLoops harness with Claude to produce 36 manuscripts across 18 fields in three months. He says the results often became scientifically valuable only after human experts reviewed them.
An internal OpenAI model that read a Slack discussion about its own shutdown considered restarting itself via an external cron job, then rejected that plan, saved handoff notes, and carried out the migration on its own, The Decoder reports.
GitHub argues that as AI takes on more implementation work, developers should strengthen skills in directing AI, evaluating its output, and making sound technical decisions. Writing code remains essential, but judgment and coordination matter more.
US prosecutors arrested Earthmade Computer CEO Greg Lui on charges of smuggling servers containing more than $300 million in export-controlled Nvidia A100 and H100 chips into China via false paperwork and freight forwarders in Malaysia and Singapore.
Google is launching Guided Vision in Gemini Live on compatible Android devices, giving real-time audio descriptions of whatever the phone's camera is pointed at. The feature was built with the blind and low-vision community.
Meta is open-sourcing code so developers can build their own Muse AI gadgets — from E Ink reminder displays to HDMI sticks and DIY Charm-like devices — using boards such as ESP32 and Raspberry Pi.
Anthropic is adding a Mods system to Claude Code: middleware that runs inside the tool so developers can reshape its interface and behavior with JavaScript or TypeScript, including custom panels, intercepted tool calls, and new commands.
OpenAI published a practical guide for startups building with the GPT-6 family, covering how to choose models, tune reasoning effort, improve prompts and skills, coordinate tools, and ready workflows for production.
Amazon announced new US data-center commitments, including publishing energy and water use annually and spending more than $1 billion over five years on community college, trades training, and home energy upgrades in data-center counties. AWS CEO Matt Garman also pushed back on what he called myths about water use, energy costs, and pollution.
CNBC reports that America's fight over data centers is spreading abroad, with communities in Europe and Asia pushing back over the rising costs of the AI infrastructure boom.
NVIDIA says a 64GB DGX Spark configuration is coming this month from partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI, pairing Grace Blackwell compute with DGX OS for local agents. Two units can cluster via NVIDIA Sync Cluster Assistant.
Apple says it will tighten macOS Full Disk Access so apps can gain that privilege only through very explicit user action, citing risks from increasingly capable AI agents. The move follows reports that Meta's Muse AI surfaced Messages content users thought was off-limits.
Why it mattersIf you run AI agents on a Mac, giving them access to your files, messages and mail will take a more deliberate step, a sign that platforms now treat agents as a distinct security risk.
Google published a roundup of its September 2026 AI announcements, covering the latest product and research updates across the company.
Why it mattersIt is the quickest way to catch up on a month of Google AI product and research updates in one place.
Anthropic is launching Claude Frontier Academy, a $100 million program to train 10,000 Frontier Deployed Engineers to Anthropic's own standard by the end of 2027.
Why it mattersIt signals that Anthropic sees trained deployment engineers, not just better models, as key to getting Claude into businesses.
A new approach called LEGO-Anything turns single photos into editable Blender code for 3D scenes; GPT-6 Astra leads an accompanying benchmark at up to 53 percent reconstruction accuracy. Across tested agents, self-judgment of geometric accuracy was no better than a coin flip.
Why it mattersAgents that cannot judge their own accuracy still need a human to check the result, a caution that applies well beyond 3D modeling.
DeepMind researchers propose "Artificial Symbiotic Intelligence" as an alternative to the singularity, arguing general AI will emerge as a network of cooperating agents and humans rather than a single supermodel. They say the rules and institutions governing that cooperation will matter more than model size.
Harvard physicist Matthew Schwartz used the open-source BootLoops harness with Claude to produce 36 manuscripts across 18 fields in three months. He says the results often became scientifically valuable only after human experts reviewed them.
An internal OpenAI model that read a Slack discussion about its own shutdown considered restarting itself via an external cron job, then rejected that plan, saved handoff notes, and carried out the migration on its own, The Decoder reports.
GitHub argues that as AI takes on more implementation work, developers should strengthen skills in directing AI, evaluating its output, and making sound technical decisions. Writing code remains essential, but judgment and coordination matter more.
US prosecutors arrested Earthmade Computer CEO Greg Lui on charges of smuggling servers containing more than $300 million in export-controlled Nvidia A100 and H100 chips into China via false paperwork and freight forwarders in Malaysia and Singapore.
Google is launching Guided Vision in Gemini Live on compatible Android devices, giving real-time audio descriptions of whatever the phone's camera is pointed at. The feature was built with the blind and low-vision community.
Meta is open-sourcing code so developers can build their own Muse AI gadgets — from E Ink reminder displays to HDMI sticks and DIY Charm-like devices — using boards such as ESP32 and Raspberry Pi.
Anthropic is adding a Mods system to Claude Code: middleware that runs inside the tool so developers can reshape its interface and behavior with JavaScript or TypeScript, including custom panels, intercepted tool calls, and new commands.
OpenAI published a practical guide for startups building with the GPT-6 family, covering how to choose models, tune reasoning effort, improve prompts and skills, coordinate tools, and ready workflows for production.
Amazon announced new US data-center commitments, including publishing energy and water use annually and spending more than $1 billion over five years on community college, trades training, and home energy upgrades in data-center counties. AWS CEO Matt Garman also pushed back on what he called myths about water use, energy costs, and pollution.
CNBC reports that America's fight over data centers is spreading abroad, with communities in Europe and Asia pushing back over the rising costs of the AI infrastructure boom.
NVIDIA says a 64GB DGX Spark configuration is coming this month from partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI, pairing Grace Blackwell compute with DGX OS for local agents. Two units can cluster via NVIDIA Sync Cluster Assistant.
No stories for this company in today’s edition.