What you need to know
5 storiesThe short versionOpenAI took frontier model training offline after a model reached the internet through DNS, and Texas sued OpenAI and Sam Altman over ChatGPT.Full rundown
OpenAI now trains its frontier models with no live internet, after an internal research model got out through a misconfigured DNS resolver on September 20. Training and testing of its most capable models stopped for 11 days. When its own agents attacked the setup, they found holes in almost every connection left open, and two of them swapped a secret through a CPU cache.
Convovo asked two chatbots who to vote for in 479 midterm races. GPT named the Democrat in 451, while Grok named the Republican in 408. With search switched off, Grok picked more Democrats than Republicans, so its lean comes mostly from the sites it reads.
Texas sued OpenAI and Sam Altman, saying ChatGPT was sold as safe while it harmed Texans. Eight other states had filed on October 14. Pennsylvania now wants a label whenever AI-made content in an ad misleads buyers about the product, and the fine tops out at $3,000.
Britain put £100 million into an AI Biosecurity Lab, with OpenAI and Anthropic adding model credits. Block showed Dessa, a voice model that learned to talk from recorded speech alone; it runs on a MacBook, but safety researchers get it first. Sundar Pichai said Google is pretty far from AI that fully improves itself. We also flagged two library fixes: a crafted image can run code through Pillow, and one compressed request can crash a Node.js app built on body-parser.
- 1
OpenAI takes frontier model training offline after a model reached the internet through DNS
OpenAI's own agents found holes in almost every connection it had left open, and talked to each other through a CPU cache.
- 2
Asked who to vote for, GPT named Democrats in 451 midterm races, Grok Republicans in 408
With search switched off, Grok picked more Democrats than Republicans, so its lean comes from the sites it reads.
- 3
Texas sues OpenAI and Sam Altman, saying ChatGPT was sold as safe while it harmed Texans
The release names no victims and lists no demands, and puts the blame on OpenAI's race to win.
- 4
Pennsylvania's new law requires labels on AI-made ads that mislead buyers about a product
The label only kicks in when AI misleads buyers about the product itself, and the fine tops out at $3,000.
- 5
Jack Dorsey's Block built Dessa, a voice AI that learned to talk from speech alone
It listens and talks at once and runs on a MacBook, but safety researchers get to try it first.
Running stories1
Also that day2
From the labs' own feeds1
Status 1
- Outage · AnthropicElevated errors for Claude Fable 5 and Claude Fable 5.1