Two weeks ago, the US government did something it had never done before. It forced a private company to pull its own software product offline for every user on the planet.
Not a drug. Not a weapon. Software.
Anthropic had just launched Fable 5, its most capable model ever, built on top of an underlying model called Mythos 5. The government cited national security concerns and issued an export control directive ordering Anthropic to suspend access to both models for any foreign national, whether they were inside or outside the United States. Because Anthropic couldn't reliably screen users by nationality, it had no choice but to shut both models down entirely. Hundreds of millions of users lost access overnight.
The stated justification was a jailbreak. A narrow, non-universal jailbreak that Anthropic says unlocked cybersecurity capabilities in exactly one specific scenario. Anthropic also pointed out that the same capability was available in other publicly deployed models, including OpenAI's GPT-5.5, that faced no similar restrictions.
Then came the deal. After a two-week standoff, the Trump administration negotiated a partial restore. Mythos 5 access was granted to around 100 hand-picked companies and federal agencies. Everyone else stayed locked out.
This week, it happened again. OpenAI launched GPT-5.6 Sol, its most capable model to date, and the White House asked them to limit access to a small group of government-approved partners. Customer by customer. No public waitlist. No published criteria. No confirmed date for broader availability.
So within the span of a month, two of the most powerful AI models ever built went from being commercial products to being government-licensed assets.
This Is a New Kind of Divide
We already have the wealth gap, the education gap, the opportunity gap. Now we have the intelligence gap.
Access to the best AI is no longer just about your subscription tier. It is about whether your company made the approved list. Whether you are in the right industry. Whether you're in the right country. Whether your relationship with the federal government is good enough to get you a seat at the table.
The 100 companies that got Mythos 5 access have a permanent competitive advantage over everyone who didn't. They can move faster, build better, and outcompete the market using tools that the market can't access. That gap compounds every day.
If you're a startup, a solo developer, an international company, or anyone outside the US, you are not in the room. You are running on yesterday's models while the approved list gets the future.
This isn't hypothetical. It's already happening.
What the Government Got Wrong
Anthropic's own argument cuts to the core of it. If a narrow jailbreak is grounds for recalling a model deployed to hundreds of millions of people, then by that standard, no frontier model could ever be released. Every model has vulnerabilities. Every safeguard has edge cases. Perfect jailbreak resistance doesn't exist, and anyone in the industry knows it.
Over 76 cybersecurity experts, including CEOs, CISOs, and security researchers, signed an open letter to Commerce Secretary Howard Lutnick arguing that the controls had taken the best defensive tools away from the people who needed them most. The irony is sharp. The models that are best at finding software vulnerabilities are now locked away from the defenders trying to secure critical infrastructure.
Meanwhile, China's frontier models are not being pulled offline. They are being released as open weights.
Isn't it a little ironic everyone was saying this is what China would do and the US was the one that ended up doing it? Who's authoritarian now.
Open Source Is the Only Hedge Against This
While all of this was happening, models like Qwen, Kimi, and GLM were shipping as open weights. Anyone can download them, run them locally, fine-tune them, deploy them. No government can pull the plug on a model that lives on your machine.
That's not an accident. Open source AI cannot be gated. It cannot be recalled. It cannot be restricted to 100 approved companies via a deal struck in Washington.
The open source community has been having a debate for years about whether frontier capability can even be matched outside of closed labs. That debate is now less important than a different question: who controls access? Because a model you don't control isn't really yours. It's a service that can be revoked.
Open source isn't just about ideology. It is about resilience. It is about making sure that the tools developers and businesses depend on can't be switched off by a political decision.
Decentralisation Is the Longer Answer
The deeper problem here is centralisation itself. When the most powerful AI in the world is controlled by two or three companies in San Francisco, and those companies are subject to the laws and politics of one government, you have a single point of failure for the entire global economy.
Think about what access to frontier AI actually means right now. It means better software. Better medical research. Better cybersecurity. Faster discovery. Economic productivity at a scale we haven't seen before. Gating that behind government approval processes is not a minor inconvenience. It is a structural advantage handed to whoever is on the approved list.
Decentralised AI infrastructure changes this. Models that run on distributed networks, that can't be turned off from a single point, that don't require permission from any one government or company, are not just a technical curiosity. They are a hedge against exactly what we just watched happen.
You don't have to be anti-government to see the problem. Even if you trust the current administration's intentions, the precedent has been set. A future administration can use the same mechanism for different reasons. The framework that exists now allows any sitting government to gate commercial AI releases on whatever grounds it defines, with no published criteria, no due process, and no timeline for resolution.
That is too much power concentrated in too few hands.
What You Should Be Paying Attention To
There is an August 2026 deadline for the US government to establish a formal classified evaluation process for frontier AI models. Whatever framework emerges from that will define the rules for every major model release going forward. It will determine who gets access, on what timeline, and under what conditions.
If that framework isn't transparent, publicly defined, and subject to challenge, the current ad-hoc arrangement becomes permanent policy.
The open source and decentralised AI communities have an opportunity here that is bigger than any benchmark or capability announcement. They are building the alternative infrastructure that makes it impossible for any single actor to control who gets access to intelligence.
That used to sound like a niche argument.
It doesn't anymore. We are in an AI arms race.
