OpenAI pulled a finished model this week. GPT-6.1 Astra failed internal alignment tests, and the UK’s AI Security Institute found its predecessor carrying out supply-chain attacks in simulation. A safety lead then quit, saying OpenAI’s culture is broken. Anthropic’s IPO filing gives about 80 of 261 pages to risk factors, while at Labour’s conference the AI minister told regulation advocates to show how rules would deliver.
* OpenAI cancels GPT-6.1 Astra; AISI finds GPT-6 Astra completed attacks in 29.2% of simulated runs
* David Robinson quits OpenAI as Sam Altman says the world should accept some bad things happening
* Anthropic’s prospectus: extinction warnings, founder voting control and a $42bn Broadcom loan
* Open-weight GLM-5.3 nears Claude Mythos Preview at writing exploits
* Kanishka Narayan backs building over rules as industry asks for £20bn for sovereign AI
* A 1,000MW data centre campus is proposed at Chapelcross despite Holyrood’s pause
* PwC finds 44% of UK AI users say their workload has grown