During roughly two weeks this month, the best cyber-defense model ever constructed was turned off, not by the supplier, but by order of the U.S. government. One of the letters was restored on Friday. After the Commerce Department informed Anthropic that it had assessed the company's safeguards as “adequate,” certain trusted partners can now resume using the Claude Mythos 5 model, which Anthropic can now redeploy to over 100 American institutions, including large companies and federal agencies. The model's public version, Fable 5, remains dark, and the letter does not mention it, though those with the inside scoop believe that Fable 5 will be back at an unspecified date.

Re-opening is the big news. The truth is the change, and who made it, who owns it and how many countries found out this month that their key to frontier technology is in one capital they can't vote in.

Walk back 2 weeks. On June 9, Anthropic released two models based on the same engine: the most powerful one it had ever sold to the general public, Fable 5, that comes with safeguards for high-risk applications, and Mythos 5, a less-restricted version for cyber defenders and critical infrastructure operators. Three days later the tides turned. The government has directed Anthropic to halt access to both models for all foreign nationals, including within the country, such as Anthropic's own foreign-national employees, due to a lack of the ability to reliably identify users based on nationality, which led to pulling both models for all users. What was to follow were daily negotiations, in which co-founder Tom Brown reportedly took the place of chief executive officer Dario Amodei to lead the negotiations. The partial relief came on the same day as rival OpenAI announced the new models to a select group of government-sanctioned partners. 2 labs and one referee.

The trigger is disputed and one of the concerns is the ambiguity. The chief executive of Amazon was reportedly alarmed when his researchers used a series of prompting to extract guarded cyberattack information from the Mythos-class version, and sources familiar with Anthropic say the company was informed about the threat within about ninety minutes of the discovery of its newest model without warning of any national-security risks. A second theme of the reporting is a suspicion, according to a single source, that a group linked to China had reached the model, something Anthropic counters by saying that it prohibits access from China outright. Even that was confused in later reports. Some officials apparently found the document that raised the alarm to be misleading because the capabilities identified also belong to a rival's model that had already been publicized.

Anthropic's defense has been steadfast and, on the merits, difficult to dismiss. It was essentially a narrow, non-universal bypass: it was an issue that asked the model to read a codebase and patch its issues, yielding a few minor, already identified weaknesses that other publicly released models could identify as well. One of the reasons it was so sharp was that it was about precedent: if one narrow jailbreak was a good enough reason to pull a model already used by hundreds of millions, every new release from every frontier lab could be the reason to pull them all. The industry responded in strong and loud harmony. Seventy-six executives, security chiefs, investors and researchers signed a letter stating that the order removed the most effective tools from defenders, created market uncertainty, and placed American leadership at risk, for no tangible security benefit, and that the Chinese models are only a few months away from the best American ones.

Not all of that suggests that the government's response was hysterical. The ability is truly disturbing. During the testing phase, this class of model has been discovered to have defects in all the major operating systems and browsers, including a 27 year old defect in OpenBSD and a 16 year defect in FFmpeg; it was reportedly able to generate working exploits on the first attempt 72 percent of the time, compared to the previous generation, which had none. Yet more than 10,000 high and critical vulnerabilities were already identified through its controlled access program. The ability to allow a defender to patch a hospital network at machine speed, is the same ability that allows an attacker to case it. That is why a government might want a hand close to the switch because it has a dual-use advantage. It's not a question of whether the worry is warranted. Who is the one who gets unplugged when the worry strikes and whether or not anyone outside of Washington gets a say.

This episode becomes a lesson in dependance here instead of a corporate spat. The order's dividing line was nationality, and that was what exploded overseas. France's president said the limits were 'bad' and that their reflex was 'purement nationaliste'; former French minister said that a country relying on others for its technology can be switched off in a flash and that AI 'must be taken like nuclear power: master it or suffer it'. Canada's prime minister used the occasion to caution against "over-reliance," stating that the misjudgment would be to take it for granted instead of diversifying. The former French prime minister said this as succinctly as anyone: “AI is now infrastructure, like electricity: Infrastructure that is controlled by others is infrastructure that others can switch off. In no uncertain terms, the European Commission's president said Europe cannot rely on others for the technologies that power its hospitals, grids and services.

It would be a misreading to take this as a European narrative. The exposure is more extensive and it affects the countries that are not yet able to develop their own. Earlier, banks and officials from the Indian finance ministry met to discuss the potential risks this technology poses, banks in India were also given access to this technology first and a request from the Chinese side was rejected. The math is stark for governments around the developing world: a resource that they do not have the capacity to replicate, did not create themselves and are increasingly requiring is being rationed by a foreign government in 90 minutes. There is no strategic autonomy when the thing you rely on goes dark without warning. The sense of self-reliance in the technologies now residing within the banking, energy and defense sectors is not nostalgia or protectionism it's simply a bill coming due: this month.

The commercial side brings into focus. Anthropic recently closed a funding round that valued the company at around $965 billion, the top among rivals valued at nearly $852 billion, and has confidentially filed to go public after its revenue run rate hit around $47 billion from around $10 billion a year ago. Just a few days away from becoming a publicly listed company, it is negotiating the product roadmap with the very government that previously indicated in its government watchlist that it was a supply chain threat, and against which the company is suing, purportedly for refusing to allow its products to be used in fully autonomous weapons and domestic surveillance. At the same time they are opponents in the court and collaborators at the table of negotiations.

But the order could be doing the opposite of what it set out to do. It was designed to prevent competitors from obtaining the finest American technology. The initial impact was in the first few weeks when there was a stampede toward the alternatives it had been intended to slow. The shares of one of the top Chinese labs rose by over 30 per cent as a new open-source release, another Chinese developer completed a record high round valued at close to $7.4 billion, multiple labs slashed prices, and demand for Chinese models reportedly surpassed American models on some platforms. The tokens of decentralized AI projects also rose. The truth behind it is that controlling the AI models is much more difficult than controlling the chips they reside on: AI models can leak just like open weights and open source can be run on a laptop, which means tools designed to choke hardware don't translate to software. The only way to measure the fallout, then, is not in some tidy slice of a pie, but in the increased gains of the companies it was supposed to box out; that's how markets read the ban.

The drama is removed and what remains is a brand-new operating system for frontier AI that's being built on the fly. The Commerce letter establishes a licensing program that will determine which companies, listed in an appendix, are allowed to use the most capable models, and officials have rejoiced at how quickly they have been able to act to ensure America does not lag behind while protected. There are two sides to speed. The framework is being constructed in the absence of any clear statute, allies are not engaged, and many consumers foreign governments, companies, the general public are unsure when (or if) they'll see access, and officials in Europe are openly frustrated by their new reliance on Washington for access decisions. The more fundamental issue is whether it is the actual "new capability" that makes the difference, or merely the existence of a "sensitive capability," as almost any frontier model would under the more liberal interpretation of the term, and the licensing net could be too wide to contain.

It's better to be partially open than blacked out, for Anthropic. However, once your product becomes strategic infrastructure, the government becomes the roadmap owner. That is the tough lesson of the past two weeks and it won't remain confined to any individual firm or nation. The switch is real, it works and the world now knows where it lives. For all of those who stand without a door in that room, allied capitals, central banks, engineers of countries not yet a few years into a model of their own, and users who have simply awoken to find a tool missing, the instructions are the same ones as a French politician gave on the day the lights went out. Construct it; diversify from it, or surrender the right to use it to someone else. There is no option 4.