Morning digest: Texas sues OpenAI, UK AI security, and sold-out AWS

Morning digest for Saturday, Oct 24. Texas sued OpenAI and Sam Altman on Friday, and one of the things the state is asking for is a court order that would stop OpenAI from releasing any new AI model to Texas users until an independent evaluator, approved by the court, has checked its safety guardrails and signed off. Then more from the last day: the UK's new money for AI security, independent scores for Kimi K3.1 Pro, Amazon saying its cloud capacity is mostly sold out through 2027, Tim Sweeney against Steam's AI labels, Meta buying an AI analytics startup and a Codex command-line update.

1. Texas Attorney General Ken Paxton sued OpenAI and Sam Altman, saying ChatGPT is unsafe for kids and that OpenAI misled users about it. The 67-page lawsuit was filed Friday, Oct 23, in a state district court in Collin County, north of Dallas. It names several OpenAI companies plus Altman personally, and it relies on two Texas laws: the Deceptive Trade Practices Act, a consumer protection law, and the SCOPE Act (Securing Children Online through Parental Empowerment), which limits what online services can do with minors' data and requires tools for parents. The filing opens with a line Altman said in 2015: "AI will probably most likely lead to the end of the world, but in the meantime, there'll be great companies." It points to violent cases, including the 2025 Florida State University shooting, where the state says ChatGPT explained to the shooter how to operate the Glock handgun he used. Texas wants the court to stop OpenAI from offering ChatGPT to minors in Texas and from offering it without age verification, to block new model releases in Texas unless "an independent third-party evaluator, approved by the Court, has reviewed its safety guardrails and approved its release," and to order civil penalties of $10,000 per violation plus the return of profits. "I will fight to protect Texans from the harms of AI and ensure that powerful technology companies are not above the law," Paxton said. Paxton, the Republican nominee for the US Senate, has previously sued or settled with Meta, TikTok, Snapchat and Roblox. An OpenAI spokesperson told The Hill the company works to improve ChatGPT's safeguards and wants to work with Texas and other states on industry-wide AI policies. What to watch: these are only the state's claims so far, and a judge would have to agree before any of the requested limits apply. The model-release condition is the one to follow, because it would put a court-approved reviewer in front of every new OpenAI model for one state's users. (Texas Attorney General, Lawsuit PDF, Texas Tribune, Courthouse News, Bloomberg Law, The Hill, @KenPaxtonTX on X)

2. The UK set out a package of AI money on Saturday, led by a £100m lab to defend against AI-assisted biological threats. Chancellor John Healey announced it ahead of next week's Budget. The biggest pieces: a £165m National AI Biosecurity Initiative, including £100m for a new AI Biosecurity Lab that the government calls "world-first," backed by up to £200m in model credits from OpenAI and Anthropic plus partnerships with Google DeepMind and Basecamp Research. There's also £120m more for the AI Security Institute, the government's AI safety testing body, £60m for a "National Cyber Shield" that uses AI agents to defend systems, and £70m for a new center that tracks software security holes in real time. On the home-grown side: £60m more for the Sovereign AI fund, now including physical and industrial AI, £18m for a chip design academy in Edinburgh, a £15m AI Challenge Prize whose first round targets health, and £10m a year to study AI's effect on jobs. "I understand and share the concerns of working people, so I am setting out a distinctively British approach to AI - one that doesn't leave workers behind," Healey said. What to watch: the biosecurity lab is the new idea here, and it leans heavily on model credits from OpenAI and Anthropic, so watch who runs it and what it publishes. (GOV.UK, GOV.UK: AI Biosecurity Lab, BBC, The Guardian, @KanishkaNarayan on X)

3. Update: Kimi K3.1 Pro's scores have now been checked independently, and it ranks fourth overall on Artificial Analysis. When Moonshot AI released K3.1 Pro on Friday morning, the scores came from the company. Artificial Analysis, which runs its own tests, gave it 55 on its Intelligence Index, up from 44 for Kimi K3 and behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). It scored 67% on Terminal-Bench 4.0, a test of coding agents working in a terminal, the highest of any model it has tested, and its rate of made-up answers fell from 53% to 44%. Pricing is $2 per million input tokens and $10 per million output tokens (20 cents for cached input), 33% less than K3. On Artificial Analysis's cybersecurity test it scored 64%, the best of any openly available model, with no safety refusals. Moonshot hasn't published the downloadable weights yet; it says they're due by Nov 4. What to weigh: early votes on Artificial Analysis's 3D test suggest it will rank lower there (Kimi K3 was 12th of 22), so if you make 3D or game assets, test it on your own work before switching. (Artificial Analysis on X, Artificial Analysis: cyber, Artificial Analysis: 3D)

4. Amazon CEO Andy Jassy says Amazon Web Services' big blocks of computing capacity are largely sold out for 2027, with a good chunk of 2028 gone too. "2027 is largely sold out. We always leave a certain amount for on demand and our inference service bedrock, or EC2, but largely sold out all the big chunks. A good chunk of 2028 is sold out," Jassy said on an a16z podcast posted Friday. Amazon expects to spend about $220 billion this year as it pushes into AI, and Jassy said its AI business reached a $25 billion yearly revenue pace in three years. Amazon reports quarterly results on Oct 29. When you'll feel it: if the biggest cloud is booked by large buyers this far out, smaller customers are likely to keep seeing waitlists and tight limits on the newest AI chips, and less pressure on prices to drop. (Reuters via Yahoo Finance, a16z on X)

5. Epic Games CEO Tim Sweeney called Steam's required AI disclosures "abhorrent." Valve's store asks developers to describe any generative AI used to make their game's content, and shows that to players. Replying on X Friday morning (7:32 AM PT) to Fal's Matt Workman, who called the disclosure premise "ABSURD in 2026," Sweeney wrote: "Forcing developers to write a political statement to be used against them by the hate mob they're helping organize is abhorrent." Valve's guidelines for the rule say: "Many artists and creators are rightfully concerned about exploitation and misuse." It comes as Steam Next Fest had more AI-assisted demos than ever and right after Epic loosened its own AI MegaJam to say entrants can use "as little or as much AI as you want." What this means for you: if you sell games on Steam, the disclosure rule hasn't changed, so keep a simple log of which assets and code had AI help. (GamesRadar+, @TimSweeneyEpic on X)

6. Meta is buying Sundial, an AI data analysis startup, to build tools and agents for its own data work. Sundial was founded in 2021 by former Meta product design vice president Julie Zhuo and Chandra Narayanan, and it automates steps like cleaning data and writing database queries. Zhuo said its customers included OpenAI and Gamma. Axios reported the deal first and Meta confirmed it: the team joins a new Meta unit for data analytics tools and agents under chief data officer Alex Schultz, and Sundial will stop serving outside customers once the deal closes. The price wasn't disclosed. What this means for you: if your team uses Sundial, start planning a move now, since it won't keep serving outside customers after the deal closes. (Axios, The Verge, @joulee on X)

7. OpenAI shipped Codex command-line version 0.164.0 with saved prompt presets. The update, released Friday, adds ten numbered /presets slots for prompts and slash commands you reuse, lets you set the model, reasoning effort and permissions in one line with /model and /permissions, shows image previews in terminals that support them, and hides credentials during some remote commands. OpenAI's Tibo Sottiaux also said switching reasoning level partway through a conversation no longer throws away the cache, so it shouldn't slow things down or eat extra usage the way it used to. Try it: update with npm install -g @openai/codex@0.164.0 and save your most-used prompt to a preset slot. (Codex changelog, @thsottiaux on X)