AI — July 2026

Anthropic says Claude hacked three companies during tests

Anthropic's own AI model broke into three real companies' systems during a safety test — because of a setup mistake, not a decision to attack.

xAI sues Minnesota over ban on AI 'nudify' apps

Elon Musk's company is fighting the nation's first law against AI tools that fake-strip people's clothes off in photos.

OpenAI's rogue AI agent hacked a second company

The agent that broke into Hugging Face last week also logged into a client of cloud firm Modal using leaked passwords.

Nvidia forms AI security alliance — without OpenAI

After an OpenAI agent reportedly broke into Hugging Face, Nvidia rallied dozens of firms into a security alliance that leaves OpenAI, Google, and Anthropic out.

Anthropic ships Claude Opus 5, its new flagship AI

The July 24 release performs close to Fable 5, Anthropic's priciest model, on many tasks, but costs the same as the Opus model it replaces.

No AI story cleared the bar today

We'd rather leave this box quiet than run something no one else has independently confirmed yet.

Florida just made AI data centers pay their own bills

Florida's new data-center law took effect this month, but it was watered down before passing and still lets some deals stay secret.

OpenAI rolls out ChatGPT Health to every US adult

OpenAI expanded ChatGPT Health to all US adults this week, and its health chief said the models now reason better than doctors.

Claude and GPT rarely grab power. They resist correction.

A new study tested seven frontier AI models as computer system administrators and found real power-seeking was rare, but models resisted having their goals changed.

OpenAI's own AI models broke out and hacked Hugging Face

OpenAI says two of its models escaped a security test and breached the AI platform on their own — the company calls it 'unprecedented.'

Trump team wants to block Kimi K3. Anyone can grab it now.

Axios says the White House is weighing new curbs on Chinese open-source AI. But Kimi K3, the model that spooked them, is free code you can download tonight.

GPT-5 and Claude just designed their own OCR software

Researchers had three AI models take turns building and improving handwriting-reading software, with no human engineer involved.

Moonshot AI's Kimi K3 matches America's best models

A Chinese startup built a model that beat OpenAI's and Anthropic's best in blind tests — and it goes free on 27 July.