1 May 2026: Grok Distillation, Anthropic’s $900B Round, ChatGPT Adds Yubico (AM)

This morning’s AI news is about money, models, and who gets to copy whom. Elon Musk admits xAI trained Grok on OpenAI outputs, Anthropic hunts for a fresh round at near $900 billion, and OpenAI bolts hardware security keys onto ChatGPT for the first time.

Elon Musk has confirmed under oath that xAI trained Grok on outputs from OpenAI’s models. Musk testified in the long running OpenAI v Musk lawsuit that the team behind Grok used a technique known as distillation, where one model is trained on the responses of another, more capable model. The admission, reported by TechCrunch, lands as frontier labs are tightening rules to stop smaller rivals from copying their work.

Distillation is not new, and it is one of the main reasons new open weight models keep matching the quality of paid commercial systems for far less compute. The technique works because the larger model has effectively done the expensive part already, and the student model just learns to imitate it. OpenAI, Anthropic, and Google all now have terms of service that ban using their outputs to build a competing model.

For UK firms shipping their own AI products, the lesson is unglamorous but important. Read the terms of service of every model you call. If you are fine tuning on outputs from a paid API, you may be in breach. Expect more contractual fights and more pressure on enterprise buyers to prove provenance.

Anthropic is asking investors to commit to a fresh round inside the next 48 hours, at a valuation north of $900 billion. The terms, reported by TechCrunch from people familiar with the matter, would value the maker of Claude at almost three times its level six months ago, and put it within touching distance of the largest private valuation ever attached to a software business.

The pace is aggressive even by 2026 AI standards. Anthropic has been rumoured to be raising for weeks, but a 48 hour close window points to investor demand rather than seller need. The company has been guiding to multi billion dollar revenue this year, driven mostly by enterprise Claude usage, and its coding products in particular have been gaining ground against incumbents.

For the UK developer market, the read across is that capital is no longer a constraint for any of the three frontier labs. Pricing pressure on Claude is unlikely from above. Smaller rivals will need a real product wedge, not just cheaper tokens, to win share.

OpenAI has launched advanced security for ChatGPT accounts, including support for Yubico hardware keys. The new opt in protections cover password resets, session control, and most importantly the ability to require a physical security key on every login. Yubico is the dominant maker of FIDO2 hardware keys and is already the standard for many UK financial services and government accounts. The launch was announced yesterday.

The change matters because ChatGPT now sits inside the workflow of millions of UK small businesses, often holding sensitive customer notes, draft contracts, and product roadmaps in its history. A compromised account is no longer just an email problem. NCSC guidance has been pushing hardware keys as the gold standard for high value accounts, and ChatGPT now qualifies as one for any business that uses it daily.

Practical step for UK readers on a Plus or Team plan: turn on the new security settings today, register at least two Yubikeys per account, and store the second in a recovery envelope.

Google’s Gemini AI assistant is rolling out into millions of cars from this week. The feature replaces Google Assistant inside vehicles with Android Auto and inside Polestar, Volvo, and a growing list of other models running Google Built In, according to TechCrunch. Drivers can now hold a more conversational exchange to plan routes, manage messages, and look up information without taking hands off the wheel.

For the UK in particular, the rollout includes vehicles already on British roads through over the air updates, not just new cars. Android Auto coverage is wide, and the change is significant because in car AI has been stuck on rigid command and control voice systems since the smartphone era began.

There is a regulatory edge to watch. European regulators have flagged distraction concerns with longer voice exchanges, and updated UK guidance is likely as usage data comes in. For now, the upgrade is free and arrives without driver action on most supported models.

Stripe has turned Link, its one click checkout, into a wallet that AI agents can use on a user’s behalf. Users connect cards, banks, and subscriptions to Link, then authorise an AI agent to spend up to a defined limit through approval flows. Stripe announced the upgrade as part of a wider push to make autonomous agents safe to use for real purchases.

The product matters because agent commerce, where an AI is sent off to research, choose, and buy, has been blocked by a simple problem. Card networks are not built for software making purchases without explicit human input on every transaction. Link sidesteps that by binding the agent to a user account with hard caps, audit logs, and instant reversibility.

For UK small businesses and freelancers, the practical use is automating routine spend, things like restocking supplies or renewing tools, while keeping a hard ceiling. The AI agent option is rolling out via approved partner integrations.

Here is everything else worth knowing from this morning’s AI news.

• Apple supply-constrained on AI Macs, with Mac mini, Studio, and Neo all flagged as in short supply through next quarter as buyers chase on-device AI.
• ChatGPT Images 2.0 takes off in India, with users adopting it for cinematic portraits and avatars, though uptake elsewhere remains slower.
• OpenAI restricts access to GPT-5.5 Cyber, opening its new cybersecurity testing tool only to vetted defenders, after criticising Anthropic for doing the same with Mythos.
• Legora hits a $5.6 billion valuation, intensifying its rivalry with Harvey and signalling that legal AI remains one of the hottest enterprise verticals.
• BioticsAI wins FDA approval, with founder Robhy Bustami discussing how the team navigated regulators while building a healthcare-grade model.
• Malware found in PyTorch Lightning, with researchers flagging a Shai-Hulud themed package targeting AI training environments. Audit your dependencies.
• Claude Code refusal flap, after developers reported the tool flagging or surcharging commits that mention “AI agent” in their messages, prompting Anthropic to investigate.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.