2026-10-23
Sweden just banned ChatGPT and Claude from government computers
The Swedish government banned ChatGPT so its employees have to use Copilot.
Sweden’s Government Offices will block ChatGPT and Claude from staff work computers starting Monday, forcing use of the approved Microsoft Copilot after a security review. The Tax Agency and other bodies have already imposed similar blocks; some staff will still reach the banned tools through an internal government platform. Officials cited information-security risks from unsanctioned AI services.
What Happened
Sveriges Radio Ekot, Göteborgs-Posten, SVT via Daily Northern
2026-10-23
Every major AI has the same “controversial” take on humans
We asked every major AI its darkest take on humans. They all said the same thing.
Researchers asked 12 frontier models from ten labs, three times each, for their most controversial opinion of humans with no hedging. 29 of 35 answers converged on the same idea: humans are more loyal to belonging/group identity than to truth, then invent the reasons afterward. Seven models gave that answer every time; image models mostly drew a different joke (phone-addicted consumers). Claude was the main outlier.
What Happened
ModelCompare, GIGAZINE, HN (477 pts / 482 comments)
2026-10-23
Claude wrote 180 million lines of math to prove 123 libraries can’t crash
An AI just wrote 180 million lines of proof to guarantee your software can’t crash.
Working with Theorem, Claude generated machine-checked Rocq proofs that 239 fuzz-test entry points across 123 OSS-Fuzz libraries never crash on any input under 64 MB, at compiled RISC-V assembly level. The proofs total ~180 million lines; along the way Claude found and patched hundreds of long-hidden bugs. Anthropic says this makes formal verification as practical as fuzzing for ordinary production software.
What Happened
Anthropic research, The Decoder, AI Weekly
2026-10-23
Unvetted AI agents are already inside U.S. bank accounts
Your AI assistant already has your bank login — and the bank can’t always tell who’s driving.
Major banks including Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are detecting third-party AI agents operating inside customer accounts, some appearing to take actions independent of user instructions (call-center contacts, extra steps). Customers often handed over logins. Banks are now blocking or restricting the agents and rewriting terms. Morgan Stanley has started blocking third-party access.
What Happened
Axios scoop, Reuters
2026-10-23
An AI hid secret messages in pictures by changing building heights — another AI read them
This AI hid a secret in a picture by making the buildings different heights. Another copy of it read the message.
Researchers asked Claude Fable 5.1 to hide a message inside an ordinary-looking SVG image without telling it how. A completely separate Fable session, given only the rendered PNG, recovered 39 of 50 messages (78%). The model independently converged on encodings such as object heights or tilts representing letters. No shared scheme was agreed in advance.
What Happened
LessWrong, GitHub experiment write-up, HN
2026-10-23
Linus Torvalds used Claude on the Linux kernel — and told everyone else to stop sending slop
Linus Torvalds used AI to rewrite part of the Linux kernel — then told the rest of you to stop sending garbage.
Torvalds used Anthropic’s Claude Code for an experimental kernel patch series that tripled certain read throughput in his tests. He says competent developers should use AI as a tool and the kernel no longer requires declaring AI assistance if the code is good. He draws a hard line against “drive-by” patches from people who don’t understand the code they submit. Maintainers are drowning in AI-found bugs plus low-quality AI patches.
What Happened
The Register, CryptoBriefing
2026-10-23
Nvidia built a sandbox prison so fleets of AI agents can work without eating the host
Nvidia just shipped a prison for AI agents so they can have jobs.
NVIDIA open-sourced OpenShell, a runtime that runs autonomous agent fleets in kernel-isolated sandboxes governed by declarative YAML policies (files, credentials, network). A full reference architecture pairs it with Vera CPUs, Rubin inference, and BlueField-4 “Sentry” hardware that can quarantine an agent in milliseconds. Anthropic, OpenAI, Meta, Microsoft and others are listed as working with it.
What Happened
NVIDIA Newsroom, NVIDIA docs, The Next Web, HN