The US government’s export control order against Anthropic’s most advanced AI models was not a national security measure, it was the visible breaking point of a structural conflict between Washington’s need to militarize frontier AI and the companies building it, a conflict that neither side has a roadmap for resolving and that will define the geopolitics of artificial intelligence for the next decade.
The Day Washington Switched Off Its Own AI
On June 12, the US Commerce Department issued an order requiring Anthropic to immediately restrict access to its two most advanced AI models, Mythos 5 and Fable 5, to any foreign national, whether inside or outside the United States. The order cited national security concerns and gave the company no meaningful implementation window. Anthropic’s response was immediate and revealed more about the state of US AI governance than any policy document has managed to: it disabled both models for everyone, because there was no technical mechanism for verifying a user’s nationality in real time.
The US government issued a national security directive, and the enforcement mechanism it chose broke the product it was trying to protect. Hundreds of millions of users lost access. Foreign researchers at American universities, H-1B visa holders at US companies, international collaborators who had built entire workflows around these models — all cut off overnight, not because they had done anything wrong, but because the order was not technically implementable without a complete shutdown. The Commerce Department reserved the right in its letter to Anthropic to “reevaluate the decisions made in this letter” if circumstances changed. That sentence, buried in bureaucratic language, is the most honest thing anyone said during the entire episode. Nobody had a plan. They were building the enforcement architecture in real time, on top of civilian technology that was never designed to accommodate it.
How Anthropic and Washington Got Here
Anthropic is not a typical technology company and its relationship with the US government is not a typical tech-government relationship. The company was founded in 2021 by Dario Amodei, Daniela Amodei, and colleagues who left OpenAI specifically over concerns about the pace of AI development and the adequacy of safety measures being applied to it. It has been more explicit than almost any other frontier lab about the risks of the technology it is building, and it has translated those concerns into contractual limits on how its models can be used.
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
Those limits are what triggered the current conflict. The Pentagon designated Anthropic a supply-chain risk earlier this year after the company refused to allow its models to be used for mass domestic surveillance and fully autonomous weapons systems. Anthropic’s position was not commercial, it was ethical. The applications were harmful enough that no contract was worth enabling them. The Pentagon’s position was that it needed AI for all lawful purposes and that a private company did not get to define what those purposes were. Anthropic sued the administration. The export control order followed months of that legal and political escalation.
The immediate technical trigger had a detail most coverage buried. It was Amazon CEO Andy Jassy personally who alerted the White House to the vulnerability, not an anonymous researcher. That matters because Amazon is simultaneously Anthropic’s largest investor, its primary cloud host, and a direct competitor building its own frontier AI model. Whether the alert reflected genuine security concern or competitive calculation has not been answered publicly, but the entity that triggered the export control order had significant financial interests in the outcome.
Anthropic disputed not just the severity but the basic characterisation of the finding. Some researchers who examined the reports described the alleged vulnerability as closer to an ordinary coding prompt than a genuine jailbreak, and roughly 100 cybersecurity experts signed an open letter arguing the export ban was causing more harm to defenders than the vulnerability ever could. The government disagreed. Within days, both models were blocked globally. The two sides were not just disagreeing about policy — they were disagreeing about whether the underlying threat was real.
The Line Anthropic Drew and the Government That Would Not Accept It
The mainstream coverage framed the Anthropic episode as a dispute about a specific jailbreak and a specific export control order. That framing is accurate at the level of events and completely wrong at the level of meaning. What the episode revealed is a structural conflict that has been building since frontier AI became a national security asset, and that NSPM-11 — the memo Trump signed on June 5 requiring agencies to terminate contracts with AI companies that repeatedly limit government use of their technology, has now been made explicit in policy.
The conflict is straightforward to describe and genuinely difficult to resolve. The US government needs the most advanced AI models for national security purposes: intelligence analysis, cyber operations, autonomous weapons targeting, and the broader AI capability race with China. The companies building those models have, in several cases, placed contractual limits on exactly those uses, not because they are hostile to American interests but because their founders genuinely believe those applications create catastrophic risks. The government’s response has been to treat those limits as a procurement problem to be overcome rather than a policy question to be debated.
Both sides are correct about the thing they are most certain about, and that is what makes this conflict so difficult. The US government is correct that it cannot afford to fall behind China in military AI. Chinese development is advancing rapidly, with significantly less regulatory friction and no equivalent of Anthropic’s self-imposed usage restrictions. DeepSeek produced models of serious capability at a fraction of the cost of American equivalents. The window for the US to maintain decisive advantage in frontier AI is real and narrowing, and the government’s urgency is not manufactured.
Anthropic is also correct that autonomous weapons systems and mass domestic surveillance are applications with consequences serious enough to justify refusal regardless of compensation. The company’s own statement acknowledged that making any AI model fully robust to jailbreaks is probably impossible and warned about the potential for a universal jailbreak that could unblock an entire class of harmful behaviors simultaneously. These are not commercial objections dressed in ethical language. They are genuine risk assessments from the people who know the technology most intimately, and dismissing them because they are inconvenient does not make them wrong.
The problem is that neither side prevails in resolving the conflict. A government that overrides Anthropic’s safety limits and forces its models into weapons applications without meaningful safeguards does not solve the jailbreak problem, it removes the company’s incentive to keep working on it. A company that refuses all government engagement eventually loses the contracts, data access, and institutional relationships that fund frontier research at the required scale. The current dynamic is pushing both sides toward that outcome faster than either appears to have calculated.
There is a geopolitical dimension that the domestic framing consistently misses. When the US government disabled Anthropic’s models for all foreign nationals, it did not just disrupt Anthropic’s user base. It demonstrated to every government, company, and research institution outside the United States that American AI infrastructure can be switched off for political reasons with no notice and no recourse. Sridhar Vembu, co-founder of Zoho, one of India’s largest software companies, said publicly that the episode proved technology is the ultimate weapon and that national sovereignty is now inseparable from technological independence. That reaction, replicated across Asia, Europe, and the Global South, is accelerating investment in non-American AI alternatives at exactly the moment American export controls were designed to prevent Chinese alternatives from gaining ground. The policy is producing the outcome it was designed to prevent.
The Scenarios
Base case (~60%): Following the pattern of every previous regulatory fight Washington has had with a technology sector it needed but could not fully control, the conflict institutionalizes slowly and becomes the new normal. NSPM-11 creates the framework for terminating contracts with non-compliant companies. Anthropic and OpenAI, both approaching IPOs and increasingly dependent on government revenue, negotiate their safety limits downward incrementally, each compromise justified by the circumstances that produced it, without any public deliberation about the cumulative direction. The erosion is invisible because it happens one carve-out at a time.
Downside case (~25%): The signals are already visible: the government has shown willingness to act faster than it can think, and the companies have shown willingness to resist longer than is commercially rational. A major AI-enabled incident triggers a crisis before any statutory framework exists to govern the response. The government will want to demonstrate control it has not built, the companies will face liability for risks they were trying to prevent, and the debate that should have happened in 2025 will instead happen reactively, under pressure, with a specific catastrophe as the framing.
Upside case (~15%): No comparable statutory framework for a strategically critical technology has ever been passed before a crisis forced the issue, that precedent is not encouraging, but it is not inevitable either. Congress uses the current NDAA cycle to establish what the government can and cannot require AI companies to enable, with democratic accountability and binding limits. It is the least likely scenario. It is also the only one that resolves the conflict rather than deferring it.
The Deferral Dressed as a Deal
The export controls were lifted less than three weeks after they were imposed, after Anthropic implemented new safeguards and agreed to inform the government of malicious activity and work collaboratively on protocols for future models. The resolution looked like a pragmatic compromise. It was a deferral dressed as a resolution, and the difference matters enormously.
The underlying disagreement about what the government can require AI companies to enable has not been resolved. It has been paused while both sides prepare for the next round. NSPM-11 is still in effect. The Pentagon’s supply-chain risk designation is still contested in court. The Commerce Department’s reservation of the right to reimpose controls at its discretion is still in the letter.
Watch for whether Congress includes any binding statutory framework for government AI procurement in this year’s National Defense Authorization Act. If it does not, and the current drafts suggest it will not — the framework being built around America’s most strategically important technology is entirely executive, improvised in real time, and reversible by the next administration or the next export control order. That is not a framework. It is a series of decisions waiting to collide.
The US will not win the AI competition with China by controlling its own companies into compliance. It will win it, if it wins it, by building the institutional trust that makes American AI worth depending on. Right now it is doing the opposite, and the world is noticing.
