An AI sign is seen at the World Artificial Intelligence Conference in Shanghai.Aly Song/Reuters
Jonah Prousky is a freelance writer and PhD student at MIT.
The U.S.-China AI race has taken on a new tenor in recent weeks, driven by growing concerns about its national security implications. The Five Eyes intelligence alliance, for example, released an alarming statement late last month calling for a “whole-of-society response” to counter novel AI cyber threats. “The timeline is not years,” it reads, “it is months.”
This change in tone was brought by Anthropic’s latest class of Mythos models, which excel at discovering and exploiting software vulnerabilities. During testing, Mythos found previously unknown bugs in every major operating system and web browser, as well as security flaws in classified U.S. systems. Anthropic restricted access to Mythos, an obvious potential cyber risk, to a cadre of tech giants and governments (including Canada’s) under Project Glasswing, a collaborative effort to “secure the world’s most important software.”
The saga deepened recently when Anthropic released Fable 5, a Mythos-class model with built-in safeguards. After reports that some users found a way to jailbreak the model, the U.S. government issued Anthropic an export control directive to suspend foreign nationals from accessing Fable 5 or Mythos 5 without a licence. To comply, Anthropic removed access to both models for all users. Partial access to Mythos has since been restored, with Fable publicly available this month, but the continued involvement of foreign nations in Project Glasswing now seems uncertain.
Opinion: Locked out of Anthropic’s Mythos, Canada must fight for AI access in USMCA talks
This has put middle powers like Canada in an incredibly awkward position. In some ways, these countries look helpless, given that the U.S. government’s directive signals that it alone will decide how widely frontier models can be distributed.
On the other hand, the U.S. doesn’t want its allies turning to Chinese companies such as DeepSeek, which are closing in on American-made LLMs. So, as was the case during the Cold War, the need for international cooperation and safety standards could create unique opportunities for middle powers to exert some influence over AI policy, especially if they band together.
Canada will be forced to walk a complex, expensive and uncomfortable tightrope. It will likely leverage the most powerful American-made models, such Mythos or OpenAI’s GPT-5.5 for national security purposes. There’s a narrow window for countries to do this while China plays catch up. But this could mean building infrastructure around a model for which the U.S. government has an “off switch.”
This tension, between capability and sovereignty, is inescapable for now. Countries can, of course, build around open-weight models onshore, but they’ll risk being quickly outmoded by the likes of Anthropic, OpenAI, and DeepSeek. Or, they can partner with U.S. firms and hope their access to American-made models doesn’t become a bargaining chip for the Trump administration, like the Gordie Howe bridge.
Recognizing that Canada will never achieve full AI sovereignty – it will probably never produce a frontier model or manufacture cutting-edge chips – the country needs to figure out what strategic dependence on the U.S. should look like. Neither Canada nor the U.S. would benefit from less cooperation on AI and national security. But the promise of mutual benefit hasn’t stopped the Trump administration from squeezing Canada on trade. The lesson is that cooperation with the U.S. cannot rely on friendship or fealty. Canada needs leverage.
Leverage means finding competitive advantages along the AI supply chain so that the country can become a partner the U.S. can’t ignore. Lots of middle powers are doing this. Taiwan is the obvious example, but there are many others. The Netherlands is home to ASML, the largest supplier to the semiconductor industry. Britain is a world leader in AI security. Canada has several advantages, namely, a highly educated workforce, AI research centres and abundant energy. But we also have a major productivity problem that has created a gap between American and Canadian firms on AI adoption.
The government’s “AI for All” strategy, much like its British and Australian equivalents, is a good attempt to bridge that gap. But this wakeup call from the U.S. suggests that middle powers need to think even bigger. For example, though it may be impossible for any one nation to produce foundation models that rival the U.S.’s or China’s, a united EU and Canada might have a shot – France-based Mistral AI, for example, has produced some promising open-source models.
Barring that, we’re stuck, as Prime Minister Mark Carney put it at Davos earlier this year, “between hegemons and hyperscalers.” Canadians deeply mistrust the U.S. and AI. But for now, our future prospects – our national security and economy – depend on both. The challenge for Canada is not to escape this uncomfortable reality, but to find enough leverage within it to shape our future.