Even by the standards of AI development, where each month often feels like a year, this has been quite the week. Back in April, Anthropic unveiled Claude Mythos, a frontier model that was supposed to be exceptionally capable of finding and exploiting software vulnerabilities, which led the company to delay its public release, which is a remarkable measure from a commercial AI lab. Instead of a public launch, Mythos was quietly deployed through Project Glasswing, giving access to the model to different tech companies and stakeholders to use Mythos defensively to patch critical software. Then, on June 9, Anthropic released Claude Fable 5, a Mythos-class model with enough guardrails bolted on to make a public launch plausible. We enjoyed a few days of the model, and the AI side of the Internet went crazy with its capabilities; I used it for a few minor tasks and it struck me as being an incredibly good model. Then the US government pulled the rug out three days later, issuing an emergency export control directive blocking access for all foreign nationals on the basis of a suspected jailbreak. Anthropic removed all access to Fable as enforcing the order would have been impossible otherwise.
The net result was that Anthropic, a company that had literally published an essay the day after Fable 5 launched arguing that governments should have the power to block dangerous AI models, found itself on the receiving end of exactly that power. You really could not make this up. And it gets better, because the essay in question didn’t just ask for the power; it asked for a specific kind of power, one exercised “as part of a statutory process that is transparent, fair, clear, and grounded in technical facts.” What it got instead was a directive delivered with no specifics, on national security grounds, that didn’t bother to engage with the technical facts at all.
I’ve been rather surprised by the relative lack of reaction outside of AI research to this news. I think that it heralds a new era of AI regulation, but most importantly, it also signals the start of what could become an AI Cold War. Once nation states become involved in AI model deployment, we are living in an entirely different world, and we should all be paying attention, as this could have immense repercussions for the future.
Dangerous models?
The underlying policy issue with the imposition of export restrictions is that there is an assumption that models can be used for nefarious purposes. One can be extremely suspicious of AI hype, and one could even believe that AI is nothing more than a glorified auto-complete, but it is clear that governments are starting to think of models as potentially dangerous under some circumstances. Whatever you think about the hype cycle, that shift in official posture is now a fact on the ground, and it operates regardless of where you sit on the question of whether any of this is “real” intelligence.
Super-intelligence can be defined as an AI that surpasses human capabilities in one area, or across the board. It could be argued that when it comes to finding and exploiting code vulnerabilities, AI may have reached better-than-human capabilities, even if it doesn’t cross the intelligence threshold. Anthropic’s own documentation shows that Mythos Preview discovered thousands of high-severity vulnerabilities across every major operating system and browser, which is exactly the sort of claim that makes governments pay attention. And there is a type of dual-use trap that runs through all of this: the capability that lets you secure critical infrastructure is precisely the capability that lets you attack it. There is no “defensive vulnerability discovery” toggle separate from the offensive one; it’s the same skill pointed in a different direction. This makes it much easier for governments to act in restricting access to models on security grounds, because the security case writes itself.
What makes this particular episode so instructive is how thin the actual trigger appears to have been. By Anthropic’s account, the supposed jailbreak amounted to asking the model to read a codebase and fix its flaws, the kind of thing security researchers do every day, producing a handful of minor, already-known vulnerabilities that other publicly available models can find without any jailbreak at all. If that is the bar for invoking national security export controls, then the bar is somewhere around ankle height. The lesson is not that Fable is uniquely dangerous; it’s that once “AI as potential weapon” becomes the governing frame, the evidentiary threshold for acting collapses. You no longer need to prove catastrophe. You need only gesture at the category.
But regardless of the Fable case, we should really start considering the implications of models that can perform actions that could be dangerous approaching super-intelligence capabilities, such as bioweapons, cybersecurity, and terrorism.
An AI Cold War
If we find ourselves in a situation in which advanced models are considered as assets that should not be trusted in the hands of rivals, we may be walking into a world in which models become nationalised assets (unthinkable), or, more plausibly, highly regulated resources subject to harsh export restrictions. We have, of course, seen this film before: the chip war has been running for years, with successive controls on advanced semiconductors and the equipment to make them. The novelty here is that the export control has been ramped-up considerably. It is no longer just the GPUs that are strategic assets; it is the models trained on them. The compute was always the obvious chokepoint. Now the weights themselves are treated as munitions.
And note just how far the Fable directive reaches, it blocks a hostile state from using the model, but it also blocks all foreign nationals, including Anthropic’s own foreign-national employees, from touching it, whether they are in San Francisco or Stockholm. A model becomes something that only citizens of one country may lawfully use. That is a genuinely strange object to introduce into the world, and it is hard to imagine other states watching this and concluding anything other than “we need our own.” Expect a renewed enthusiasm for sovereign AI, national champions, and domestically controlled frontier labs, because if the most capable models are going to be gated by nationality, then nationality is precisely what every government will now optimise for.
So far, only the USA and China have access to the top frontier models, with Europe a far-away third (regardless of all of the Mistral Le Chaton Gros memes). Some countries such as the UK may have had some good image models (such as Stable Diffusion), but are nowhere near the top of the line. If we start thinking of AI as a national asset, countries may want to start developing their own capabilities fast. Or we could encounter a future of research blocks, much like the Cold War (I propose Oceania and Eurasia, just for the heck of it).
The above would lead to a situation of considerable inequality. A handful of countries would have access to the best models while everyone else makes do with whatever is permitted to cross the border, and the gap between the two would compound exactly as fast as the models improve. The first casualty of a national-security framing is openness. Open weights are, almost by definition, incompatible with a regime in which models are strategic assets, because you cannot export-control something you have released to the world. For those of us who have spent a long time arguing for open source, the Commons, and a right to tinker, this is the part that should worry us most. It is one thing to lose access to a proprietary model behind a paywall; it is quite another to have open release reframed as a national security liability, with the entire ethos of open development treated as a loophole to be closed. The chip war gave us a fragmented supply chain, and the new model war threatens to give us a fragmented world, in which the model you can use depends on your passport.
Concluding
Imagine a world in which access to the best models is predicated on where you live, in which open weights models are forbidden, and where AI development has become a matter of national security. On the one hand, this could be defensible if we genuinely believe that models could be extremely damaging in the wrong hands. On the other hand, it would entrench an inequality that no amount of “AI for everyone” marketing will paper over, and it would do so while handing governments a discretionary power whose exercisemay not be transparent at all.,
That is the bit worth discussing in detail. Anthropic wanted governments to have the authority to block dangerous deployments, and they were right that someone, eventually, will need that authority. But they wanted it constrained, evidence-based, and proceduralised. What they received was a reminder that you do not get to choose how a power is used once you have argued it into existence; you only get to choose whether it exists. Be careful what you wish for…
The reason so few people outside AI research have reacted is, I suspect, that this still looks like an industry squabble, in other words, a commercial lab in conflict with a regulator. It is not. It is the moment the model stopped being a product and became a strategic asset, and that is a one-way door. Whatever happens to Fable 5 specifically, whether access is quietly restored next week or not, the precedent is now set. Models can be export-controlled on national security grounds, on contested evidence, overnight. We should all be paying attention. I know that the Cold War analogy is a bit over the top, but the structure is there: rival blocs, controlled technology, a scramble for sovereignty, and a great many countries about to discover they are non-aligned whether they like it or not.
On the meantime, maybe the best winning move is not to play. How about a nice game of chess?
0 Comments