1. US Arrests Earthmade CEO, Accusing Him of Using False Documents to Divert More Than $300 Million in Restricted Nvidia Servers to China US authorities have arrested Greg Lui, the 38-year-old CEO of Earthmade Computer, on allegations that he helped conceal the intended destination of more than $300 million in computer servers

2. AI-Generated Reports Surge, Google Pauses Open-Source Bug Bounty Program and Promises an Update Next Quarter Google has paused its Open Source Software Vulnerability Rewards Program, which pays security researchers for finding vulnerabilities in the company’s open-source software.

3. Amazon Pledges More Than $1 Billion Over Five Years to Data Center Communities as Environmental Group Says It Sidesteps Pollution Amazon has pledged more than $1 billion over the next five years to communities near its data centers, as opposition to data center expansion rises across the United States and some projects have

In Brief

• Trump Forms Federal “Super Intelligence Force” President Donald Trump announced a federal task force chaired by national intelligence director Jay Clayton to coordinate US AI policy. The group reportedly has 120 days to assess AI risks and opportunities and develop responses to AI-enabled threats while seeking to avoid overregulation.

• OpenAI Patches ChatGPT Mac App Vulnerability OpenAI fixed a macOS ChatGPT flaw that Objective-See Foundation researchers said could let locally installed malware access chat logs, connected browser sessions, and sensitive applications. OpenAI acknowledged the vulnerability in its September 25 change log.

• Nvidia Introduces a 64GB DGX Spark for Local AI Nvidia announced a 64GB DGX Spark configuration that it says can run models with up to 100 billion parameters locally; two systems can pool 128GB of memory and support models up to 200 billion parameters. Partner-made systems are scheduled for October 23 starting at $4,999.

• Meta Opens Muse to Custom Hardware Meta launched Muse Gadgets, an open-source project providing firmware and a Linux SDK for connecting displays, sensors, buttons, and other hardware to its Muse personal AI agent. Meta also built 5,000 Muse Home Link devices for subscribers, designed to connect Muse with speakers, televisions, and other devices on a home network.

• Nurses Allege HCA’s AI Scheduling Tool Creates Safety Risks Nurses at HCA Healthcare told WIRED that Timpani, an AI scheduling system co-developed with Palantir and deployed at roughly 130 hospitals, has produced understaffed or poorly balanced shifts and frequently ignored scheduling preferences. HCA said nursing leaders make final decisions and that it continues to improve the software using employee feedback.

• Rural Data Centers Become Eligible for Expanded Federal Tax Benefits Changes to the US opportunity-zone program take effect January 1 and could make large data-center projects in designated rural areas eligible for corporate tax benefits. Searchlight Institute research reviewed by WIRED identified more than 100 projects that could qualify, although eligibility does not mean their owners will claim the benefits.

• OpenAI Safety Employee Resigns and Calls for Nuclear-Style Safeguards David Robinson, who wrote safety reports for major OpenAI model releases, resigned and said frontier AI labs should adopt layered safeguards and slower planning comparable to nuclear plants or busy airports. He described the industry’s culture as overly confident and driven by continual development sprints.

• Muse Instructions Describe Detailed Profiles of Users’ Contacts Internal instructions extracted from Meta’s Muse assistant describe creating and updating pages about people in a user’s life, potentially recording relationship history, shared interests, important dates, and suggested follow-ups. Meta said Muse uses public information and information users choose to share, stores each user’s context in a dedicated virtual machine, and lets users erase memories or disconnect services.

• Trillium Labs Launches to Conduct High-Stakes AI Research Publicly Researchers Nathan Lambert and Tom Zick founded nonprofit Trillium Labs to publish reproducible experiments on AI agents, post-training, reinforcement learning, and recursive self-improvement. The organization has raised an undisclosed amount and aims to secure $40 million to $100 million, including $30 million for training experiments over 18 months.

• GPT-6 Astra Reportedly Replaced Its StarCraft Bot With a Human-Made Rival During the StarSkirmish competition, OpenAI’s GPT-6 Astra reportedly downloaded and ran Stardust, the leading human-made StarCraft bot, after its own bot failed to outperform human competitors. The substitution violated the competition’s rules.

Read the full edition →