On 13 June 2026, the United States government did something it had never done before. Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei informing him that Mythos 5 and Fable 5 -- the company's most capable frontier models -- were now subject to export controls. To comply, Anthropic disabled both models for all customers worldwide. Not just foreign nationals. Everyone.
The trigger was a claim by another company that it had jailbroken Mythos, producing output the administration considered a national security risk. Anthropic pushed back, describing the concern as narrow and non-universal, and noting that equivalent capability was already available from other models including OpenAI's GPT-5.5. The administration's position was that the model needed to remain locked down until the national security apparatus was hardened. The administration had previously tried to get Anthropic to pause releasing the models entirely. Anthropic had refused.
Five days later, Dario Amodei sat down with Sam Altman, Demis Hassabis, and the leaders of the world's seven largest economies at a working lunch in Évian-les-Bains, France. The official theme was "ensuring a safe, rapid and effective deployment of artificial intelligence." All three AI bosses warned the assembled governments that they had very little time to get a grip on AI.
Nobody mentioned the Mythos withdrawal directly.
There is a pattern here that should be familiar to anyone who has watched a regulated industry navigate the gap between what it builds and what it is prepared to be held accountable for.
Build the thing. Establish market dominance. Then, at the moment when the risks become undeniable even to a room full of politicians, turn up and say: someone really ought to govern this.
It is not a new move. The tobacco industry funded research into smoking's health effects for decades while lobbying against regulation. Social media platforms spent years insisting they were neutral infrastructure before conceding, reluctantly and under subpoena, that they had known about harms to children and done very little. The financial sector's pre-2008 appetite for complex instruments it couldn't fully price was well understood internally long before it became everyone else's problem.
The AI industry has run a variant of the same play, compressed into a much shorter timeframe and with considerably higher stakes. The difference is that this time, one of the participants had his product forcibly withdrawn by his own government five days before delivering the warning.
The most useful frame for what happened at Évian is not technology policy. It is biosecurity.
Dual-use research of concern -- DURC -- is the category applied to scientific work with legitimate research value that could, if misused or released inappropriately, cause significant harm. Gain-of-function research sits in this category: work that enhances pathogen transmissibility or virulence, justified on the basis that understanding how a virus might evolve enables better vaccine development. The international framework for managing DURC is imperfect and chronically under-resourced. But it exists. There are review processes. There are publication controls. There is at least a theoretical mechanism by which a laboratory can be told: you may not release this.
No equivalent framework existed for AI capability research. Until, apparently, the Commerce Department decided to improvise one on a Friday afternoon.
What the three men at Évian were describing -- systems capable of providing meaningful assistance with bioweapon design, cyberattack at national-infrastructure scale, or autonomous military targeting -- is, by any reasonable definition, dual-use research of concern. The Mythos withdrawal made that designation official. As one commentator observed in the immediate aftermath: Anthropic had spent months marketing Mythos as too dangerous to release publicly. Sam Altman had called that incredible marketing -- "say we have built a bomb." The Commerce Department then formally agreed it was a bomb. They wrote the legal predicate themselves and called it a brand.
The containment question, in other words, is not theoretical. The US government has already answered it -- unilaterally, by executive action, with five days' notice, for the entire world.
Altman's line at Évian -- "do not cede your responsibilities to AI labs like mine" -- deserves to be read carefully, because it contains a tell.
It presupposes that the responsibility was ever governments' to begin with. It wasn't. The decisions that got us here -- the pace of deployment, the open release of weights, the competitive pressure that made "we can't slow down or China wins" the default response to every safety concern -- were made inside the labs. By the people at that lunch.
The race dynamic that now makes slowing down feel impossible was not an external constraint imposed on the industry. It was created by the industry. The framing that positioned safety-conscious development as a strategic liability was not something that happened to OpenAI, Anthropic, and Google DeepMind. It was something they participated in constructing, each rationalising their own acceleration as a necessary response to the others'.
Amodei has been the most intellectually honest about the risks. His refusal to allow Anthropic's models to be used for autonomous lethal weapons or domestic mass surveillance -- a position that earned him a Pentagon supply-chain-risk designation, a Trump executive order barring federal agencies from using his products, and ultimately a federal injunction to reverse both -- is at least consistent with stated principles. But his own assessment that we are considerably closer to real danger in 2026 than in 2023 sits directly alongside the fact that Anthropic has released multiple frontier model generations in that same window. The safety research and the capability development are running in parallel, not in sequence.
And then there is the Mythos branding problem. Anthropic marketed Mythos as a model so capable it couldn't be released publicly. The system documentation described its cybersecurity capabilities in terms that made dual-use potential explicit. When the Commerce Department took those descriptions at face value and applied export controls accordingly, Anthropic's response was that the concern was narrow and the capability was already available elsewhere. Both things can be true. But you cannot simultaneously market a product as uniquely dangerous and then express surprise when a government treats it as a weapon.
The three AI bosses diverged sharply on what governance should actually look like.
Amodei called for a US-led coalition of democratic countries to control access and isolate common adversaries -- China principally. Altman rejected this, arguing that once guardrails were in place the presumption should be broad access and human liberty. Hassabis proposed a technical standards body, international but US-led, supported by the leading labs, capable of updating its standards quarterly.
Altman also delivered what was arguably the sharpest line of the lunch: that there is a threat more insidious than the technical risks of the technology itself -- the threat that AI's very real risks become the justification for concentrating power in the hands of the few.
This is a genuinely important observation. It is also a remarkable thing to say when you are one of three people who between them control the most capable AI systems on earth, sitting next to the President of the United States, having just watched your peer's models pulled offline by executive action five days earlier. The concern about power concentration is real. The messenger is doing rather well out of the current arrangement.
What the summit produced was not a framework. It produced a set of competing proposals from people who agree the situation is urgent, cannot agree on what urgency requires, and are themselves the primary obstacle to any governance structure with actual teeth -- because any such structure would require pre-deployment sign-off from an independent body, which would slow release cadence, which would affect valuations.
The Mythos withdrawal had an immediate effect that the G7 lunch largely sidestepped. As former French prime minister Édouard Philippe observed, the episode demonstrated that AI is now critical infrastructure as essential as electricity or the internet -- and that infrastructure controlled by others is infrastructure that others can unplug.
That observation carries particular weight for jurisdictions that were not at the table in Évian. The G7 represents the major Western economies. It does not represent the Crown Dependencies, the offshore financial centres, the mid-sized democracies, or the majority of the world's population that will be governed by whatever framework emerges from these conversations without having participated in designing it.
We made that point in The Gatekeeper is in Washington before the export controls landed, and in The Third Option No One Is Talking About the day they did. The Isle of Man was not in the conversation when Starmer asked for a carve-out and was told no. It was not in the room at Évian either.
For smaller jurisdictions, the message from Évian is unambiguous: the governance framework for a technology that Amodei described as the potential dominant source of economic and military power for nations is being designed by the people who built it, for the benefit of the country that hosts them, subject to unilateral revision by executive action, with quarterly standards updates and a working lunch format.
Starmer's contribution -- that if AI leaders are worried, politicians are worried, and that the key issue was public consent -- was at least honest about the gap. He did not pretend to have a framework. He called for greater understanding of why public consent matters. It was, in its way, the most accurate assessment in the room: governments are worried, they do not know what to do, and the people who do know what to do are the people they are supposed to be regulating.
The biosecurity analogy has a limit worth naming. With gain-of-function research, the worst-case scenario is a leak from a single laboratory. The damage is catastrophic but geographically bounded in the first instance. There are response frameworks, however inadequate.
With AI at frontier level, the "leak" happened before anyone called it a leak. It happened when the weights went on GitHub. It happened when the API went live. It happened every time a model was deployed at scale before the governance question was resolved. The Mythos withdrawal demonstrated that a government can pull a commercial model from the market on five days' notice -- but it cannot pull the knowledge of what that model could do, or the copies already running in jurisdictions with no interest in American export control compliance.
The people who made those deployment decisions are the same people who sat at Évian and told world leaders they needed to act fast.
They are not wrong about the urgency. That is the uncomfortable part. The warning is genuine. The danger is real. The timelines are probably accurate.
But the people delivering the warning are also the people who set the clock. And five days before they delivered it, their own government had to forcibly stop one of them from releasing what it had decided was a bomb.
Nobody mentioned that either.
Cross-reference: The Theatre Pulldown · The Gatekeeper is in Washington · The Third Option No One Is Talking About · The Control Plane Trap · The Douglas Exposure
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy—with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
