COLUMN: American AI under seizure — when Washington cuts access to its own brains

5:21 PM, a Friday evening, a letter that changes everything

On Friday, June 12, 2026, at 5:21 PM Eastern Time, a letter signed by Commerce Secretary Howard Lutnick landed in the inbox of Anthropic, one of the world's most important artificial intelligence companies. The message was brief, imperative, offering no detailed explanation: the US government, invoking its "national security authorities," was ordering the company to immediately suspend all access to its two most advanced models — Fable 5 and Mythos 5 — for any "foreign national, whether inside or outside the United States, including Anthropic's foreign employees."

The problem was this: Anthropic cannot distinguish US nationals from foreigners in real time among its hundreds of millions of users worldwide. The only way to comply with this directive was to cut access for all users, everywhere in the world. That is what Anthropic did. Within hours of receiving the letter, Fable 5 — launched three days earlier to enormous fanfare — and Mythos 5 were offline for everyone. Everywhere. Without warning.

Global stupefaction and the American paradox

The reaction across the technology world was immediate and stunned. Not only did users in Europe, Asia, and the Middle East lose access to the models, but Anthropic's own engineers — the people who had built these tools — found themselves locked out if they held foreign citizenship. One international trade law expert summed up the absurdity: "The ammunition is in the building and the people who made it are not allowed to look at it." America had just placed its own artificial brains under seizure.

This is one of the most self-destructive decisions I have seen from a Western government in a long time. In trying to protect its technology from China, Washington handed China and its open-source models the best publicity imaginable: proof that the United States does not trust its own tools. The contradiction is breathtaking.

Mythos 5: the model that frightens its own creators

Mythos 5 is the most advanced model ever developed by Anthropic. Its distinguishing feature is exceptional capability in the domain of offensive cybersecurity — it can identify software vulnerabilities, propose exploitation methods, and reason about complex security flaws. For precisely this reason, Anthropic had decided not to make it publicly available: since its launch in April 2026, access to Mythos had been limited to approximately 200 organizations in 15 countries, selected to test vulnerabilities and contribute to the model's security. These organizations were cybersecurity firms, research laboratories, and government agencies.

Fable 5 is the public version of Mythos, with built-in guardrails that block responses in high-risk domains — offensive cybersecurity, biology, chemistry, model distillation. This was the model Anthropic had launched on June 9, 2026 for the general public, after spending months building these protections. Three days after its launch, it was cut.

The jailbreak claim: the central question

The trigger for the government directive was the claim by Amazon researchers to have found a method for bypassing Fable 5's guardrails — a "jailbreak." Amazon CEO Andy Jassy had reported this discovery to Treasury Secretary Scott Bessent, who escalated it up the chain of command. Former national AI coordinator David Sacks stated the government had asked Anthropic to fix the problem or pull the model — and that Anthropic had refused, arguing the jailbreak was "narrow and not representative of a significant risk." That was when Lutnick signed the directive.

The government's version is that this is a critical jailbreak on a model capable of massive cyberattacks. Anthropic's version is that it is a narrow, non-universal jailbreak that the entire industry encounters regularly and that does not justify a global shutdown. Someone is lying or exaggerating. What is certain is that the decision was made without transparency, without clear procedure, without any avenue for appeal. That is the real problem.

Europe in panic over digital sovereignty

The global shutdown of Fable 5 and Mythos 5 sent an immediate shockwave through Europe. Governments, companies, and universities that had been using these models for applications ranging from medical research to cybersecurity to machine translation found themselves with no access overnight. Even more alarming: the American decision confirmed a latent European fear that had been building for years — that dependence on American technology platforms exposes European actors to unilateral decisions from Washington over which they have no say.

As Fortune reported, the shutdown of Anthropic's models "caused panic in Europe over digital sovereignty" — and "joy among China's open-source AI developers." That formulation perfectly encapsulates the paradox: in trying to protect its technology, Washington reinforced the arguments of its adversaries and allies who want to reduce their reliance on American AI.

OpenAI and Google: involuntary beneficiaries

In the commercial world, the two immediate beneficiaries of Anthropic's misfortune were its direct competitors: OpenAI and Google. Anthropic customers that had integrated Fable into their systems — and suddenly found themselves without service — had to turn to alternatives. Anthropic's sales figures paradoxically increased immediately following the directive, according to TechCrunch — the effect of the involuntary publicity generated by the controversy. But this commercial irony does not mask the structural problem.

Every time the US government makes a decision this brutal and opaque about its own companies' technology, it hands China an additional argument for convincing the entire world to develop alternatives. Chinese open-source models like DeepSeek — which anyone can download and run locally, beyond any possible government control — became more attractive the instant Fable 5 was cut.

Friday evening: Anthropic tries to negotiate

According to detailed journalistic accounts from Politico, the hours immediately after receiving the directive were frantic. Anthropic's founder and CEO, Dario Amodei, was outside Washington when the letter arrived. A source close to Anthropic stated he was reachable and available, contrary to government claims. Once contacted, Amodei participated in three phone calls with senior administration officials — including national cybersecurity coordinator Sean Cairncross and Commerce Secretary Lutnick. He argued that the discovered jailbreak was narrow, specific, and did not justify a global shutdown. Government officials, according to Politico, were not convinced — they were relying on Amazon's research findings as evidence.

Anthropic therefore published both its compliance and its contestation. In a blog post, the company stated: "We received the government directive today at 5:21 PM ET. We are complying with the legal government directive and removing access to Fable 5 and Mythos 5 for all users. We apologize for this disruption to our customers. We believe there is a misunderstanding and are working to restore access as soon as possible."

The weekend of intense negotiation

Over the following weekend, intense negotiations took place between Anthropic teams and several senior administration officials. During the week of June 15, senior Anthropic representatives — researchers and security experts — physically visited the Department of Commerce and the Office of the National Cyber Director in search of a technical compromise. These meetings concluded without immediate resolution. Meanwhile, both sides were working on a "joint framework" of AI model security standards — a technical oversight regime the administration wanted in place as a precondition for lifting the restrictions.

What happened over that weekend looks like a government that acted on impulse — perhaps under pressure from Amazon trying to harm a competitor — and then found itself managing the consequences without a plan. The technical negotiations for a post-crisis security framework should have preceded the decision, not followed it. This is the Trump government: action first, thought later.

Export controls applied to AI models: a first

The June 12, 2026 directive represented a major legal first: it was the first time the US government had applied export control laws to a commercial artificial intelligence model. These laws, specifically the Export Administration Regulations (EAR), were designed to control exports of military and dual-use technologies. Consumer-grade AI models had never been treated as exportable technologies subject to control.

The Trump administration had issued an executive order in early June asking AI companies to voluntarily submit their models for government evaluation 30 days before launch. This process was described as voluntary — but the action against Anthropic showed that participation may not have been as voluntary as promised. Legal experts cited by Politico noted that the directive appeared to contradict prior Bureau of Industry and Security guidance, which had concluded that allowing foreign nationals to access cloud computing resources did not in itself constitute an "export."

The deemed export rule: the paradox of foreign engineers

The so-called "deemed export" rule is the most contested legal provision in the directive. This rule holds that showing controlled technology to a foreign national in the United States is equivalent to exporting it abroad. Applied to AI models, this meant that Anthropic's foreign-national engineers — who likely represent a substantial proportion of the technical workforce at major Silicon Valley AI companies — could no longer access the models they had themselves helped create. The legal, ethical, and operational paradox was total.

Silicon Valley was built on the world's best minds. The Indian, Chinese, Iranian, European, and African engineers who developed America's most advanced technologies are now considered security risks by the US government. This is a worldview so paranoid it turns against itself. You cannot build technological superiority by excluding the world's best talent. It is a fundamental contradiction.

June 26: conditional authorization for 100 companies

On June 26, 2026 — two weeks after the initial crisis — Commerce Secretary Lutnick sent Anthropic a letter authorizing the company to give access to Mythos 5 to a group of approximately 100 carefully selected companies and federal agencies. The word used in the letter to describe them was "trusted partners." This partial authorization was progress, but left Fable 5 — the public model — still offline for the general public.

According to Bloomberg, unconfirmed reports from June 27 indicated that Anthropic and the Trump administration were "on track toward a deal" that would allow restrictions on both models to be lifted. Talks centered on a technical security framework — model verification procedures, standards for reporting jailbreaks, government oversight mechanisms — that both sides appeared ready to accept.

A mandatory oversight regime for AI: the implications

This unfolding resolution reveals what the Trump administration was really after: not a permanent ban on advanced models, but the establishment of a mandatory oversight regime for frontier AI models prior to commercialization. The June directive and the ensuing crisis functioned as a show of force — demonstrating to Anthropic and the entire industry that the government had the means to impose its will, and that it was in everyone's interest to cooperate in shaping the rules rather than have them brutally imposed.

I understand the government's logic: a model capable of detecting and exploiting zero-day vulnerabilities in critical software is potentially a weapon. China, Russia, and Iran are desperate to access it. Guardrails are needed. But the method chosen — global shutdown without notice, without transparency, without recourse — is that of a government that has not yet learned to regulate intelligently. There is a difference between AI governance and bureaucratic violence.

The suspicion of Chinese access to Mythos

A crucial dimension of this affair is revealed by Semafor: one of the motivations for the government directive was suspicion that China-linked groups had potentially accessed the Mythos model through the testing partners program. If accurate, this would explain why the government reaction was so brutal and rapid. Mythos-class models — capable of advanced offensive cybersecurity — are precisely the type of technology Beijing seeks to acquire for its own cyberwarfare capabilities.

Anthropic denied that its discussions with the administration concerned Chinese access to Mythos — stating the government had not raised this angle during negotiations over the Fable 5 jailbreak. But it is possible these two matters were handled separately, or that classified intelligence not shared with Anthropic played a role in the decision.

DeepSeek and Chinese open source: the big winner

The decision to shut down Fable 5 and Mythos 5 handed a public relations victory to the Chinese model DeepSeek and the Chinese open-source AI ecosystem in general. The logic is straightforward: if a government can cut access to a proprietary American model overnight, actors who want reliable, uninterruptible technology not subject to Washington's whims have an additional reason to turn to open-source models they can run locally. Developers in Europe, Southeast Asia, and South America who had invested in the Anthropic ecosystem received the message: your infrastructure can be shut down without notice.

In trying to prevent China from accessing American technology, Washington handed China the most powerful argument possible for convincing the world not to depend on American technology. This is the paradox of absolutist security: by trying to control everything, you lose everything. Machiavelli would have laughed.

Europe confronting sovereign powerlessness

For Europe, the Anthropic crisis was a brutal revelation of a reality that many still refused to admit: European actors — governments, companies, universities, hospitals — that have integrated American AI tools into their critical infrastructure are potentially vulnerable to unilateral decisions from Washington. This vulnerability is not theoretical. It materialized in June 2026, for several days or weeks, for millions of users.

The political response in Europe is moving in two parallel directions: reinforcing European sovereign AI initiatives — whose budgets remain far below American and Chinese investment levels — and demanding contractual guarantees from American providers on service continuity. Both take time, and in the meantime, dependency remains the operational reality.

Global AI governance: the abyssal void

The June 2026 crisis also highlighted the total absence of an international framework for AI governance. Unlike nuclear weapons — governed by the Non-Proliferation Treaty (NPT) — or chemical weapons — regulated by the Chemical Weapons Convention — militarily capable AI is subject to no international oversight regime. A country can unilaterally decide to cut global access to a model, without international consultation, without transparent procedure, without any avenue for appeal.

We need an international treaty on dual-use AI models. Not to ban their development — that is both impossible and counterproductive. But to establish minimum standards of transparency, notification, and procedure when a government decides to impose restrictions. What happened on June 12, 2026 should never happen this way again. Never.

A nuanced statement, a difficult position

Anthropic's public response throughout this crisis was remarkably calibrated. The company complied with the directive while publicly contesting its validity. It declared: "We believe the government should have the authority to prevent dangerous deployments through a process that is transparent, fair, clear, and based on technical realities. This action does not meet those principles." This is a direct challenge, without taking the matter to court — which Anthropic apparently has no intention of doing, preferring to maintain its working relationship with Washington.

This posture reveals Anthropic's fundamental contradiction: it is a company that presents itself as the champion of safe and beneficial AI for humanity, yet depends on both private markets and government contracts. It cannot afford open war with Washington. It must negotiate, compromise after compromise, with a government that does not yet have a coherent AI policy.

Anthropic vs. the Pentagon: the broader context

The June crisis is part of a broader context of tensions between Anthropic and the US administration. According to Semafor, the Pentagon had previously characterized Anthropic as a "supply chain risk" because of the company's refusal to allow its models to be used in fully autonomous weapons systems. Anthropic maintained a policy of refusing deployments in autonomous lethal applications — an ethical position that had earned it hostility from parts of the American military establishment.

Here is Anthropic's Kafkaesque situation: refusing to let its models kill autonomously puts it at odds with the Pentagon. Leaving a potential gap in its guardrails puts it at odds with national security. Both positions are consistent with its stated ethics. But the state cannot tolerate nuance. Either you are in, or you are out.

Toward a mandatory oversight regime for frontier models

The probable resolution of the Anthropic crisis — an agreement on a technical security framework with conditional access granted to "trusted partners" — outlines the contours of a new AI governance regime in the United States. Companies developing frontier models — the most technologically advanced ones — will need to submit these models for prior government evaluation before commercialization. If a security problem is identified, the government can impose restrictions. This regime, if it stabilizes, represents a break with the near-total freedom tradition that had prevailed in American commercial AI development until now.

The stakes are considerable. If this regime is poorly designed — too restrictive, too opaque, too dependent on discretionary power — it risks slowing American innovation while driving talent and capital to less constrained jurisdictions. If instead it establishes clear, predictable, technically grounded standards, it could become a model for the rest of the world.

The technology competition with China: the real issue

Behind this entire affair lies the fundamental question of the technology competition with China. Advanced AI models are now recognized as strategic assets comparable to semiconductors, missile systems, or space technologies. China is investing massively in AI, developing its own frontier models, and actively seeking American technologies by every available means — including economic intelligence and talent recruitment.

The real Cold War of the 21st century is being fought in data centers, not in trenches. And in this war, cutting off your own tools for fear the enemy might access them is like burning your library so the barbarians cannot read it. The barbarians will find other books. You will have lost yours.

European sovereign AI: an urgent and underfunded goal

The Anthropic crisis has reinforced the arguments of those who, in Europe, have been calling for years for a sovereign AI strategy. Initiatives exist — GAIA-X for data, IPCEI programs for semiconductors, the European Commission's AI project — but investments remain far below those of the United States and China. France, Germany, the United Kingdom, and Sweden have AI centers of excellence, promising startups, and world-class universities. But no European model today approaches the capabilities of Fable 5 or Mythos 5.

The question for Europe is not to cut itself off from American models — that is a short-term illusion. It is to build its own capacity that progressively reduces asymmetric dependency, and to negotiate contractual guarantees from the United States on service continuity for critical applications. Neither is simple. Both are necessary.

Toward multilateral AI governance?

The June 2026 crisis relaunched discussions on the need for an international AI governance framework — discussions that had begun at the Bletchley AI Safety Summit in 2023 and were continuing without producing binding commitments. The UN, G7 and G20 forums, the Council of Europe — several bodies are working on this. But the pace of technological change far exceeds the tempo of multilateral diplomacy.

I am not naive about global AI governance. National interests, the technological arms race, great-power mistrust — all of this makes a binding international treaty on dual-use AI extremely difficult. But the June 12, 2026 crisis reminds us why we must try anyway. Unilateral decisions harm everyone.

AI startup funding: an uncertainty signal for investors

The seizure of Fable 5 and Mythos 5 affected not only Anthropic — it sent a shockwave through the entire AI startup ecosystem. The venture capital funds that are investing heavily in AI companies reacted with alarm to the US government's demonstration of its capacity for unilateral intervention. If a company as established as Anthropic — with billions in funding and government contracts — can have its products shut down overnight, what security can smaller, more vulnerable startups have?

According to sources cited by Bloomberg, several AI startup fundraising discussions were paused in the days following the Anthropic crisis, with investors waiting to see how the situation resolved before committing further. European funds began incorporating specific clauses on American regulatory risk into their valuations of American startups. This uncertainty is costly for innovation: it drives capital toward more predictable regulatory environments.

Universities and academic research: model access compromised

A less-covered aspect of the crisis is its impact on academic artificial intelligence research. Thousands of researchers at universities worldwide were using Anthropic's models — notably Mythos 5 via its API version — for work in natural language processing, bioinformatics, computational social science, and other fields. The sudden cut in access disrupted research projects, blocked ongoing analyses, and forced some laboratories to urgently rebuild their work pipelines around alternative models. Doctoral theses depending on experimental results from these models were affected.

Scientific research needs stability and predictability. When the US government can cut access to fundamental research tools within hours, without warning, without appeal, it compromises not only one company's commercial interests but an entire global scientific ecosystem. This is collateral damage the directive's authors appear not to have calculated.

The philosophical debate underlying the Anthropic crisis

Behind the legal, geopolitical, and commercial dimensions of the Anthropic crisis lies a fundamental philosophical debate: who should decide what an artificial intelligence can do, and for whom? The US government's answer is implicit in the June 12, 2026 directive: the state, in the name of national security, has the final word. Anthropic's answer is different: companies that develop AI tools carry the responsibility — and the technical competence — to decide on appropriate safeguards, with transparent government oversight grounded in objective criteria.

These two positions reflect two deep worldviews. The first is paternalistic and state-centric: the government knows better than companies what is dangerous and must therefore be able to intervene sovereignly. The second is liberal and technocratic: innovative companies, supported by external evaluators, are best positioned to assess the risks of their own technologies and implement appropriate protections. The problem is that both views have strong arguments — and the reality of the situation likely demands a synthesis.

AI safeguards: between real security and regulatory theater

One of the central technical questions in this crisis is the reliability of the guardrails built into advanced AI models. Jailbreaks — techniques for bypassing model restrictions — are a real and documented problem across the entire AI industry. No model is perfectly resistant to every conceivable jailbreak. The question is whether a partial jailbreak on a specific scenario justifies a total shutdown of a model for all users worldwide. For Anthropic, the answer was no. For the US government, in this specific case with this specific model, the answer was yes. This technical disagreement about acceptable risk levels is at the heart of the dispute.

There is no easy answer to the question of AI safeguards. A model capable of identifying vulnerabilities in critical software can help defensive cybersecurity teams — or help malicious actors launch cyberattacks. Same technology, same two uses. This is the fundamental paradox of dual-use AI. And neither companies nor governments have yet found the right line of demarcation.

The civil aviation model: certification before commercialization

To think about an AI governance framework that balances innovation and security, experts suggest drawing inspiration from civil aviation. Commercial aircraft are not put into service without prior certification by regulatory authorities — the FAA in the United States, EASA in Europe — which verify system safety according to precise, transparent technical criteria. This process is costly and extends time to market, but it produces reliable security standards that are comprehensible to industry. An analogous framework for advanced AI models — with clear evaluation criteria, appeals processes, and international cooperation — is technically feasible and politically desirable.

Concrete proposals already exist: the US National Institute of Standards and Technology (NIST) has published an AI risk evaluation framework that could serve as a foundation. The British AI Safety Institute, created after the Bletchley Summit, is working on evaluation protocols. The European Parliament adopted the AI Act, which creates risk categories and transparency requirements. These elements exist — but they have not yet been coordinated into a coherent international system.

Allied cooperation: sharing evaluations, not just results

A specific dimension of this crisis is the impact on US allies. The American decision of June 12, 2026 was taken without prior consultation of NATO allies whose companies, governments, and universities were using Anthropic's models. This unilateralism is politically costly — it reinforces the arguments for European technological autonomy — and practically counterproductive, as it deprives allies of the information needed to protect their own infrastructure against the same risks. A cooperative approach, where security evaluations of frontier AI models are shared with allies before any restriction decision, would preserve both security and Alliance cohesion.

AI governance is an international coordination challenge we have not yet mastered. But the tools exist: multilateral treaties, standards bodies, intelligence-sharing agreements among allies. What is missing is the political will to deploy them before the next crisis. Because there will be a next crisis — possibly more serious than the one of June 2026.

What this crisis reveals about Washington and AI

The Anthropic crisis of June 2026 reveals several uncomfortable truths about the relationship between Washington and artificial intelligence. First, the US government does not yet have a coherent, transparent, and technically competent policy for managing frontier AI models. Second, it is willing to use heavy, opaque regulatory tools — export controls — to impose its will on private companies, without a clear appeals process. Third, these decisions create massive collateral damage for allies and global users, without prior consultation.

An empire beginning to fear its own tools

There is something profoundly revealing in this image: the US government, which has largely funded and enabled the development of the world's most advanced AI, is afraid to leave that AI in the hands of its own companies and their global users. This is the paradox of a technological power that acknowledges the dangerousness of its own creations — a legitimate acknowledgment in some cases — but responds with closure rather than intelligent regulation. The future of AI deserves better than this absurd dance between ambition and fear.

I close this column with the stubborn hope that humans are capable of governing their own creations with wisdom. The history of technology shows we get there — never as fast as we should, often after a few disasters, but we get there. AI deserves better than panic or indifference. It deserves our attention, our vigilance, and our collective ambition.

By Maxime Marquette, columnist

Editorial stance

I am not a journalist but a columnist and analyst. My expertise lies in observing and analysing the geopolitical, economic and strategic dynamics that shape our world. My work is to take political strategies apart, to understand global economic movements, to put the decisions of international actors in context and to offer analytical perspectives on the changes that are redefining our societies.

I do not claim the cold objectivity of traditional journalism, which stops at reporting facts. I claim analytical clarity, rigorous interpretation and a deep understanding of the complex issues that concern us all. My role is to give meaning to the facts, to place them in their historical and strategic context, and to offer a critical reading of events.

Method and sources

This text keeps the basic distinction between verified facts and interpretive analysis. The rule never changes: a factual claim is published only when a verifiable source supports it, and the sources this article actually used are listed under Sources, never here.

Primary source families used by the publication, when the story has them: official statements from governments and international institutions, public statements by political leaders, reports from intergovernmental organisations, dispatches from recognised international news agencies.

Secondary source families: specialised publications, internationally recognised news media, analyses from established research institutions, reports from industry organisations.

When an article cites statistical, economic or geopolitical data, it comes from the institutions that produce it (intergovernmental organisations, central banks, national statistics offices), and the exact institution is named under Sources.

Nature of the analysis

The analyses, interpretations and perspectives in the analytical sections of this article are a critical, contextual synthesis based on the information available, the trends observed and the expert comments cited in the sources consulted.

My role is to interpret these facts, to set them in the context of today's geopolitical and economic dynamics, and to give them a coherent meaning within the larger story of the changes shaping our time. These analyses reflect expertise built through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors.

This article describes a documented state of affairs on its publication date, not a prediction: later developments may shift its outlook. No update is promised in advance; when an article is corrected or completed, the change is dated in the text.