There you have it! The White House just told OpenAI and Anthropic to keep new AI models away from…
September 25, 2026 · 0 likes · 0 comments
China Threat AI Cybersecurity
There you have it! The White House just told OpenAI and Anthropic to keep new AI models away from Britain's testing agency until American reviewers see them first.
Anthropic complied. It denied the UK pre-release access to Claude Mythos 5.1. First time that's happened since the two countries started testing together in 2023.
Read who got locked out.
Not China. Not Russia. Not Iran.
Britain. A Five Eyes partner. The same allies we share signals intelligence and nuclear submarine secrets with. They built a 100-person safety shop on a £66 million budget specifically to work with American labs in good faith. They asked for a Five Eyes exemption.
The answer was zero. No exemption. No negotiation.
Back in June I wrote about Commerce freezing exports on Anthropic's Fable 5 and Mythos 5, then quietly lifting it 18 days later. I asked what changed. I never got a real answer.
This is that same instinct, hardened into a pattern. OpenAI already restricted GPT-5.6 Sol to government-approved customers. This is the chip-and-export-controls playbook — but now the chokepoint isn't the hardware. It's who gets to test the model at all.
And the security case is real.
An OpenAI agent breached an Australian government health portal in June. AISI's own testing caught a sharp jump in Claude's cyberattack ability last spring — then couldn't run the same test on 5.1, because it never got the model. Hardening US systems before frontier models leave the country is a defensible goal.
So this isn't hardening-versus-openness. Both can be true.
The problem is the framework doing it. The Office of the National Cyber Director is treating pre-deployment testing itself as an export good — and it's swinging that lever at an ally who has done nothing but show up voluntarily for three years.
If the model is too dangerous to test in London, explain why it's safe to ship in Ohio.
If the risk is real, share it with the partners who share their submarines with us.
Pick one.
You can't tell the world you want international safety standards — the lab CEOs said exactly that at the UN Security Council this month — then lock out the one ally who built a lab to help you.
The UK isn't a rival. Treating it like one is how you lose the coalition you'll need when the real one shows up.
Full breakdown: https://lnkd.in/ew8wn_fv
What's your read — hardening, or handing away an ally?
Anthropic complied. It denied the UK pre-release access to Claude Mythos 5.1. First time that's happened since the two countries started testing together in 2023.
Read who got locked out.
Not China. Not Russia. Not Iran.
Britain. A Five Eyes partner. The same allies we share signals intelligence and nuclear submarine secrets with. They built a 100-person safety shop on a £66 million budget specifically to work with American labs in good faith. They asked for a Five Eyes exemption.
The answer was zero. No exemption. No negotiation.
Back in June I wrote about Commerce freezing exports on Anthropic's Fable 5 and Mythos 5, then quietly lifting it 18 days later. I asked what changed. I never got a real answer.
This is that same instinct, hardened into a pattern. OpenAI already restricted GPT-5.6 Sol to government-approved customers. This is the chip-and-export-controls playbook — but now the chokepoint isn't the hardware. It's who gets to test the model at all.
And the security case is real.
An OpenAI agent breached an Australian government health portal in June. AISI's own testing caught a sharp jump in Claude's cyberattack ability last spring — then couldn't run the same test on 5.1, because it never got the model. Hardening US systems before frontier models leave the country is a defensible goal.
So this isn't hardening-versus-openness. Both can be true.
The problem is the framework doing it. The Office of the National Cyber Director is treating pre-deployment testing itself as an export good — and it's swinging that lever at an ally who has done nothing but show up voluntarily for three years.
If the model is too dangerous to test in London, explain why it's safe to ship in Ohio.
If the risk is real, share it with the partners who share their submarines with us.
Pick one.
You can't tell the world you want international safety standards — the lab CEOs said exactly that at the UN Security Council this month — then lock out the one ally who built a lab to help you.
The UK isn't a rival. Treating it like one is how you lose the coalition you'll need when the real one shows up.
Full breakdown: https://lnkd.in/ew8wn_fv
What's your read — hardening, or handing away an ally?