Is the US really pulling the plug on advanced AI models?

Two of the most advanced AI models in the world were suddenly pulled offline after the US government hit them with an export ban. The move has triggered a wave of confusion and debate: are we watching responsible regulation in action, or a political overreach that could slow down US leadership in AI?

What happened to Mythos 5 and Fable 5?

Anthropic developed a powerful new AI model called Mythos 5. Before release, the company warned that Mythos might be “too dangerous for public consumption,” especially because of its potential to uncover cybersecurity vulnerabilities. Instead of a full public rollout, they limited access to a small set of companies.

To make a safer, more widely available version, Anthropic created a tuned-down model called Fable 5. Fable 5 had guardrails that restricted certain topics, including detailed cybersecurity guidance, and was released to hundreds of millions of users.

Just days later, the US government under the Trump administration imposed export controls on both Mythos 5 and Fable 5, citing national security concerns. In response, Anthropic removed access to both models for all users, not just those outside the US.

The jailbreak that changed everything

According to reports referenced in the discussion, someone found a way to “jailbreak” Fable 5. Jailbreaking in this context means bypassing the model’s safety filters so that it behaves more like the original, more capable Mythos 5.

Once that jailbreak was demonstrated, the logic became awkward for Anthropic. The company had publicly framed Mythos as too dangerous to release, then shipped Fable as a safer alternative. If Fable could be turned back into Mythos through a jailbreak, regulators could argue that the “too dangerous” capabilities were effectively available anyway.

This is the core irony highlighted by security expert Gary McGraw: Anthropic’s own safety narrative about Mythos being uniquely dangerous made it easier for the government to justify stepping in once Fable’s protections were shown to be fragile.

Why the US government stepped in

The export ban came shortly after an executive order requiring AI companies to give the US government early access to their most advanced systems before public release. The stated goal: understand potential security risks, especially around cyberattacks and critical infrastructure.

From the administration’s perspective, if an AI model can help discover or weaponize software vulnerabilities at scale, it starts to look less like a productivity tool and more like dual-use technology—something closer to a digital weapon.

That’s why the situation is being compared to restrictions on advanced chips or military hardware. The export ban on Mythos 5 and Fable 5 is effectively treating top-tier AI models as sensitive technology that shouldn’t freely flow to foreign actors.

Is Mythos really “too dangerous” – or just overhyped?

Not everyone in the technical community buys the idea that Mythos 5 is uniquely dangerous. McGraw and others argue that Mythos is roughly comparable in power to other frontier models and that many of them can already help find cybersecurity issues with similar efficiency.

The deeper problem, he notes, is that we don’t yet know how to rigorously measure the security of AI models. Current benchmarks and leaderboards tell us something about performance on tests, but they don’t capture how these systems behave in open-ended, real-world scenarios—especially when users are actively trying to break the rules.

Some researchers, like Melanie Mitchell at the Santa Fe Institute, have argued that we need new ways to probe AI models as if they were “alien intelligences,” not just bigger versions of traditional software. Until we have those tools, claims that one model is safe and another is dangerously beyond the line are hard to validate.

Who should decide when an AI model is too risky?

This clash exposes a core governance question: who gets to decide when an AI model is too risky to deploy—companies, governments, or some independent body?

Anthropic’s co-founder Chris Olah has publicly acknowledged that AI labs operate under intense commercial, geopolitical, and personal incentives. Even with good intentions, those pressures can conflict with “doing the right thing.” That’s why he has called for outside critics—religious leaders, civil society, scholars, and governments—to scrutinize AI labs and push them toward safer behavior.

The export ban shows what happens when that outside scrutiny comes with real power. The government didn’t just criticize; it effectively hit the off switch for Anthropic’s flagship models. For users and businesses depending on those tools, that raises a new risk: if a regulator can suddenly pull the plug, how reliable is any AI platform as critical infrastructure?

AI as the new cryptography war

McGraw compares the current AI moment to the “cryptography wars” of the 1990s. Back then, the US government treated strong encryption algorithms as munitions and restricted their export. Only experts could fully understand what was allowed and what wasn’t, and the result was that non-US regions doubled down on building their own cryptographic tools.

The same pattern could repeat with AI. If the US heavily restricts its most powerful models, other regions may accelerate their own open or domestic alternatives. That could weaken US influence over AI standards and safety practices, even as it tries to maintain control.

We’re already seeing a global race in AI models and agents. For context on how fast this ecosystem is moving, you can look at roundups like recent waves of new AI models and platforms, which show just how quickly new contenders appear.

Could this trigger a brain drain in US AI?

Export controls don’t just affect where models can be used; they can also affect who wants to work on them. If top foreign researchers feel they can’t fully participate in frontier AI development in the US—or that their work could be shut down overnight—they may choose to work elsewhere.

McGraw points out that high-end AI progress depends heavily on scientific and engineering talent: PhDs, researchers, and builders at universities and labs. If the policy environment feels hostile or unpredictable, it could push that talent to other countries or to more open ecosystems.

That risk sits alongside other structural shifts in AI, like the rise of powerful agents and deep research tools from companies such as Google. For example, Google’s push into enterprise-grade agents, explained in detail in this breakdown of Gemini Deep Research and enterprise agents, shows that the competition is not just about raw models but also about who can attract and retain the best teams to build on top of them.

Two layers of irony at the heart of this clash

McGraw describes the situation as “two ironies colliding”:

First, Anthropic framed Mythos 5 as too dangerous to release, then tried to walk that back with Fable 5. Once Fable was shown to be jailbreakable, they were forced to live with their own earlier narrative.

Second, the US government is now restricting what it calls its most important AI technology from the rest of the world, echoing past attempts to lock down cryptography. History suggests that such restrictions don’t fully work and can even backfire.

Neither position is comfortable: companies don’t want to lose control over their products, and governments don’t want to be blamed for ignoring real risks. The result is a messy, very public negotiation over where the lines should be drawn.

Where AI safety and policy need to go next

Underneath the politics, there’s a real need for better science and clearer rules. We need:

• Transparent methods to measure what AI models can and can’t do, especially in security-sensitive areas.
• Shared standards for when a model crosses into “high-risk” territory.
• Processes that involve companies, governments, researchers, and the public in decisions about deployment and restrictions.

McGraw argues that we should aim for openness about how models are built, what data they’re trained on, and how we test them. That doesn’t mean open-sourcing everything, but it does mean building trust through clearer, collectively understood criteria instead of ad hoc reactions.

For now, we’re watching this play out in real time. One night, users are happily working with a cutting-edge AI assistant; minutes later, it’s gone because of a government directive. That volatility is a sign that AI has moved from experimental tech to something closer to critical infrastructure—and our governance tools haven’t fully caught up.

What this means for everyday AI users

If you’re using AI tools in your work or business, this saga is a reminder of a few practical realities:

• Access to specific models can change suddenly due to regulation or policy shifts.
• Marketing claims about “too dangerous” or “uniquely safe” models should be treated with healthy skepticism until we have better measurement tools.
• The AI tools you rely on are increasingly shaped not just by engineering choices, but by geopolitics and national security decisions.

The big question going forward is whether we can build a system where powerful AI is developed, deployed, and governed in a way that’s both safe and stable—without stalling innovation or concentrating too much power in any single set of hands.