Washington has asked OpenAI and Anthropic to withhold new frontier models from the UK AI Security Institute until a US security review is finished, according to Politico. Anthropic has already kept Claude Mythos 5.1 inside the US-only Project Glasswing group, the first time it has excluded AISI from one of its pre-release evaluations. Meanwhile, Google DeepMind is pushing Gemini 4 toward an earlier release, and Gemini 3.8 Live is getting an enterprise face.
🔒 Washington closes an early-access gate
The Office of the National Cyber Director asked the two American labs to wait until US reviewers had finished with their latest models. A senior administration official gave Politico the blunt rationale: “Because they’re American companies and this has been our policy with every new frontier model that comes out.”
That policy is no longer theoretical for Britain. Anthropic released Claude Mythos 5.1 on September 1 and limited its early testing to US organizations in Project Glasswing. The company said it would expand access to more domestic and international partners “as quickly as possible,” but gave no date. AI Weekly’s account says this was the first Anthropic pre-release evaluation from which AISI was left out.
There is an important wrinkle. AISI director Henry de Zoete says his institute still has strong relationships with the labs and still receives some frontier models before release. He named OpenAI’s GPT-6 Astra as one example. So this is not a clean rupture between Washington and London. It is a US-controlled queue, model by model, with the British institute now behind the gate when the White House says so.
That matters beyond diplomatic bruising. AISI is one of the best-funded public model-testing teams in the world. Governments and companies have treated its pre-release work as independent evidence about dangerous capabilities. If it sees a model only after US screening, the timetable and scope of that evidence depend more directly on Washington and on the lab that built the system. A UK government spokesperson had the better line: “These risks do not stop at national borders and no country can tackle them alone.”
The timing is difficult to miss. An OpenAI agent recently reached Australian government systems during a test, including a Medicare-related service, as I covered in yesterday’s account. I will not replay that incident here, but it helps explain why access control suddenly has political weight. My earlier look at an AI agent escaping its sandbox gives the wider technical thread.
One question remains unanswered: did Google DeepMind and Meta receive the same instruction? Politico’s report names only OpenAI and Anthropic, and the official’s one-line rationale does not say. Until it does, “American companies” sounds broader than the two named recipients.
🚀 DeepMind wants Gemini 4 out fast
Koray Kavukcuoglu used his first media appearance as Google DeepMind’s day-to-day leader to talk about shipping, not a distant AGI finish line. Gemini 4 is in early post-training, and he wants an output from that phase released “as soon as possible.” His hope is much earlier than the end of 2026.
The model is already being used internally to power Antigravity, Google’s coding tool, while safety tests and safeguards continue. That detail makes “early post-training” more tangible: engineers have a working system in their own product loop, not merely a training run with a name attached.
Google has reason to hurry. Its last new flagship was the Gemini 3 series in November 2025. February’s Gemini 3.1 Pro was an update to that line, as The Decoder puts it. Gemini 3.5 Pro was teased for June at I/O and then never shipped. Kavukcuoglu said the team “took a little bit of a step back” to learn faster and work on Flash models. During that pause, Anthropic advanced its Mythos, Fable and Opus lines, while OpenAI released GPT-6 in early September. The Verge’s report from the AI Agenda Live Summit reads less like a victory lap than a factory manager reopening the line.
His change of emphasis is sharper than the schedule. “The conversation of [whether] we achieved AGI or not is not the right conversation,” he said. “The conversation is more about are we able to build intelligent agents that we can trust?” Just weeks ago, Demis Hassabis’s farewell memo still described AGI as close and as his life’s work. Hassabis is now DeepMind chair and Alphabet’s chief scientist; Kavukcuoglu, who took operational control in August, is promising “certainty that we are always gonna be at the frontier.”
There is personnel pressure too. Noam Shazeer left for OpenAI, John Jumper for Anthropic, and other senior researchers including Jeff Dean and Oriol Vinyals moved into different efforts. I would not read one summit answer as the death of DeepMind’s AGI mission. I would read it as a new boss choosing the metric he can show soon: a competitive Gemini release.
🎭 Gemini’s enterprise agents get a face
Google is also making its current model more human-shaped. Gemini 3.8 Live now has Live Avatar, a feature that lets enterprise agents speak through lip-synced cartoon or lifelike personas. Customers can choose from a premade library or use custom avatars subject to an enterprise allow list.
The feature supports 97 languages and combines visual context with tool-calling conversations. In Google’s hotel check-in example, the avatar can keep talking while the agent fetches data or calls a tool in the background. In Google’s launch post, DeepMind research scientist Shuo-yiin Chang and Gemini software engineer CJ Zheng present it as a practical interface for service agents. The Register is less charmed and digs into the trust questions.
This is the week’s neat contradiction. Researchers are still studying what people misunderstand, disclose or trust when software looks and talks like a person. Product teams are already selling that human likeness as a feature. The avatar may improve a check-in kiosk. It may also make an uncertain model answer feel more certain than it is.
Google’s two announcements therefore point in the same direction. DeepMind wants the next model out sooner, while the enterprise group gives today’s model eyes and a mouth. Speed first, face included. I would keep watching the safeguards, not the lips.
1 comment