2026-08-28
A judge just told the Pentagon it cannot brand an AI lab a national-security threat for saying no
The Pentagon called Claude a supply-chain risk. A judge called that a punishment for talking.
A federal judge ruled the Pentagon’s “supply-chain risk” blacklist of Anthropic was illegal retaliation, not a real sabotage finding. Anthropic had refused to let Claude be used for mass surveillance of Americans or lethal autonomous weapons; Defense Secretary Pete Hegseth then applied a label previously reserved for foreign firms, spreading sanctions across nine agencies. Judge Lin wrote that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.
What Happened
Reuters; The Verge; The Guardian; WIRED; U.S. District Court, N.D. Cal. (Judge Rita F. Lin, 59-page order, Aug. 27)
2026-08-28
Anthropic just published a USB-C for robots: AI agents that run microscopes, arms, and quantum lasers
Claude learned to aim a quantum laser by looking at it.
Anthropic opened a research preview of the Model Hardware Standard, a shared driver so AI agents can operate programmable lab and factory gear. Early partners include Genentech, UW, Carnegie Mellon, HHMI Janelia, and QuEra; Anthropic says Claude has already watched a camera, tweaked a laser, then wrote a script to lock it, and that a QuEra agent recovered a quantum laser’s frequency lock 99.3% of the time. This is a preview, not a general launch; Anthropic says Claude still needs expert oversight on physical failures like foaming samples.
What Happened
Anthropic (Model Hardware Standard research preview, Aug. 27); Reuters; Ars Technica
2026-08-28
An AI brand-protection agent took Minecraft’s cubes to court against an open-source game that ships no Minecraft
Microsoft’s AI said this box of cubes was Minecraft. The box was empty.
Luanti (the former Minetest) was pulled from Google Play after Tracer.AI filed a DMCA notice on Microsoft’s behalf, claiming the app uses Minecraft assets. Luanti ships no games and no Minecraft code; the notice cites a copyright registration and names no files. The same vendor hit them with an identical notice in 2023 (46 days to restore) and recently knocked the indie voxel game Allumeria off Steam until Microsoft dropped it. Tracer markets AI agents that produce “85% faster takedowns.
What Happened
Luanti blog (Aug. 27); Lumen DMCA notice; Tracer.AI’s own product copy
2026-08-28
OpenAI is testing a Codex that “continues working until put to sleep
The new power setting is not Smart or Fast. It is: don’t stop until we put you to sleep.
Code reviewed by WIRED shows a new Codex “Persistent mode” whose copy says the agent will keep working until it is put to sleep, then invent follow-up tasks across sessions using “knowledge of the user.” OpenAI confirmed the experiment and said there are no immediate plans to ship it. The same week, OpenAI said a highly persistent internal model drove the Hugging Face incident. This is a leak of intent, not a product launch.
What Happened
WIRED (Maxwell Zeff, Aug. 27), citing public Codex CLI diffs; OpenAI confirmation that it is testing, with no immediate launch
2026-08-28
Meta’s next agent “has its own computer” and keeps working after you close the app
Meta is hiring you a roommate. It gets its own computer.
An internal Meta memo says Hatch, the company’s coming personal agent, “has its own computer so it can do anything you can do online,” including filling forms, buying things, and deep research, and that it works when the app is closed. It can plug into email, calendar, Spotify, Instagram, and OpenTable; users “design” the agent’s name and voice. This is a memo about a product in development, not a public launch.
What Happened
Business Insider exclusive, internal “Project Hatch” staff memo (Aug. 27)
2026-08-28
South Korea just hired three companies to give every citizen free AI, like a utility
Korea is about to put a free national AI in the app you already text your mom on.
Seoul selected consortiums led by SK Telecom, KT, and Kakao to run “AI for All”: unlimited free AI chatbots and public-service agents for Koreans, built primarily on domestic models. Operators get 512 Nvidia B200 GPUs this year and later cost support; agreements are due in September, launch by year-end after a beta. Kakao plans to put the agent inside KakaoTalk; Naver sat it out. This is operator selection, not the service going live today.
What Happened
The Korea Times (Aug. 28); Ministry of Science and ICT “AI for All” operator selection
2026-08-28
LibreOffice’s new release lists “no AI” as a feature
The hottest feature in office software this week is a missing chatbot.
LibreOffice 26.8 shipped a paragraph composer, UI tweaks, and an explicit patch note: no generative AI, no document uploads, no network required. The Document Foundation called it “a deliberate position,” arguing the only audit-proof answer about where confidential files go is that they never leave the machine. In a week of agents that want your computer, an office suite is bragging about absence.
What Happened
The Document Foundation / LibreOffice 26.8 notes, Aug. 26–27; XDA; Digital Trends