In an opinion piece in the Financial Times, Chris McGuire — senior fellow for China at the Council on Foreign Relations and former US National Security Council deputy — argues that April 2026 is an AI inflection point: US labs have built models so powerful they are deliberately not releasing them publicly. Anthropic's Claude Mythos Preview is described as the first model able to autonomously discover, chain and exploit or patch software vulnerabilities faster than nearly any human researcher, at scale. Cyber-security experts call it a "watershed event". OpenAI's upcoming model Spud will reportedly follow the same selective-release playbook.
McGuire's core argument: the current US lead over China in AI is roughly seven months, and that window is exactly how long America has to harden its digital infrastructure before Chinese cyber-weapons match US defences. China's AI ecosystem is still structurally dependent on US technology — Nvidia chips (despite sanctions), distillation from US models, and ASML lithography machines (China bought more DUV systems in 2024 than every other country combined). Close the loopholes and the lead extends to 18 months or more.
His policy prescription is maximalist: halt all AI-chip exports to China (including the less-advanced Nvidia H200), crack down on smuggling networks (he cites one case diverting 2.5 billion dollars of Nvidia servers via South-East Asia), block Chinese cloud access to export-controlled chips, prevent remote access to US AI models, and stop semiconductor manufacturing equipment exports including foreign-made kit that relies on US tech. The Cold War precedent he offers: the US blocked Soviet nuclear assistance but later shared permissive action link technology to prevent unauthorised launches. Maximise the lead, but keep talking on guardrails. Source: Financial Times, 24 April 2026, Chris McGuire.