Trump just found the worst way to regulate AI

In a sense, Anthropic CEO Dario Amodei is getting what he wanted.

Amodei has long argued that AI is becoming dangerously powerful — and thus, that regulatory restrictions on the technology are urgently needed. In an essay published last week, Amodei wrote that the release of cutting-edge AI models “should be blocked or reversed as a threat to public safety” if they fail to meet strict security standards. 

Alas, asking the current US government to assume sweeping new regulatory authorities is a bit like wishing on a monkey’s paw (or, for the zoomers in the audience, a “One Wish Willow”): Days after Amodei’s manifesto went live, Anthropic’s latest AI model went dark, on orders from Uncle Sam.

That model — known as “Mythos” in its unrestricted form and “Fable” in its heavily bounded, publicly accessible one — represents a major technical achievement. On conventional benchmarks of AI performance, it greatly outscored all of its predecessors. And during its brief public tenure, countless users marveled at its abilities. In my own tests of its journalistic skills, Fable proved 30 percent more effective than past models at inducing feelings of obsolescence and existential dread.

Anthropic initially shared Mythos exclusively with vetted public and private organizations, so that they could steel their cyber-defenses against its capabilities. Before releasing its new model to the general public, Anthropic lined it with strict safety guardrails: Fable will refuse to answer virtually any query about cybersecurity or biology (to prevent its use for hacking and bioterrorism).

The White House deemed this insufficient. On Friday, after learning that Fable contained a potential security vulnerability, the administration imposed export controls on the model — making it unlawful for Anthropic to provide Fable to any foreign national, including its own immigrant employees. In practice, this meant that Anthropic needed to take Fable offline completely (AI models still can’t scan their users’ brains and confirm their nationalities). 

In other words: Our government has claimed the power to block or take down AI models that threaten public safety.

But Amodei isn’t celebrating. And other proponents of AI safety probably shouldn’t either. 

True, the White House’s initial, laissez-faire approach to AI governance now lies in ruins. Emerging from the rubble, however, is the worst kind of regulatory regime: one governed by the executive branch’s whims (rather than clear and binding rules), the apparent technical misunderstandings of lay officials (rather than the knowledge of domain experts), and a corrupt president’s political biases (rather than the impartial dictates of law or cost-benefit analysis). 

America needs a regulatory system that mitigates AI’s risks, while facilitating its benefits — not one that enables the president to kneecap his least-favorite companies on dubious grounds. And the White House appears to be building the latter.

The case for banning Fable

At first glance, the administration’s actions might look reasonable. After all, Anthropic itself was unnerved by Mythos’s gifts for cybercrime. And even with guardrails, Fable is exceptionally powerful. On its face, it’s not implausible that the model could pose unique security challenges.

What’s more, one of Anthropic’s own investors warned the White House that Fable was vulnerable to a potential “jailbreak” — meaning, a method for circumventing the model’s safety controls.

Last Thursday, Amazon — which has a $13 billion stake in Anthropic — shared research documenting such a jailbreak with administration officials. The White House responded by reaching out to Anthropic and asking it to fix the issue. The AI firm insisted that its model was safe and that the administration was misunderstanding Amazon’s research. 

The administration therefore concluded that Anthropic was unable or unwilling to fix the problem. It then decided that imposing export controls on the model was the only way to ensure that it did not degrade America’s cybersecurity.

Fable’s security liabilities might be roughly the same as ChatGPT’s

Yet this version of events is incomplete. And upon closer scrutiny, the administration’s behavior looks less defensible. 

Specifically, there appear to be (at least) three problems with its crackdown on Fable.

First, it is plausibly rooted in a technical misunderstanding. No existing AI model is 100 percent jailbreak-proof. And the specific capabilities that Amazon identified are not unique to Fable, according to some experts. Katie Moussouris, head of the cyber security group Luta Security, reviewed a copy of Amazon’s findings and told the Financial Times that they raised no novel risks: According to Moussouris, Amazon showed that, when prompted in a certain way, Fable would identify software vulnerabilities, ostensibly to help the user shore up their defenses. But many frontier models, including OpenAI’s GPT 5.5, will provide the same service.

For its part, Anthropic says it subjected Fable to thousands of hours of testing — by independent organizations and the US government — to ensure that it contains no universal jailbreak, which is to say, “a method that can very broadly bypass the model’s safeguards, unblocking a wide range of cyber capabilities.” But it insists that the kind of narrow jailbreaks identified by Amazon are impossible to fully preempt. 

If this is right, then the administration’s targeting of Fable would be selective and capricious. 

The Fable crackdown may be politically motivated

Second, there is good reason to believe that the administration’s heavy-handed actions were informed by Anthropic’s refusal to curry its favor. 

Earlier this year, Anthropic and President Donald Trump’s Defense Department got into a conflict after the AI firm refused to approve the use of its models for mass surveillance and fully autonomous weapons systems. The Pentagon responded by declaring Anthropic a “supply chain risk” — a designation that would restrict the capacity of government contractors to do business with it. 

This measure was legally dubious and transparently disingenuous; essentially, the administration was asserting that Anthropic’s AI was structurally unsafe for government work, even as it continued using that very AI for government work. The policy’s plain intention was to punish a company that had insisted on contractual terms that the administration did not like.

This precedent alone offers grounds for doubting the White House’s impartiality in imposing export controls on Fable. And the fact that the administration is cozy with two of Anthropic’s top competitors — OpenAI and Elon Musk’s xAI — adds further cause for skepticism.

But the best evidence for the administration’s bad faith comes from its own explanations of its actions. In an interview with Axios, a “source familiar with the administration’s thinking” said that Anthropic’s difficulties partly reflected its inability to “communicate effectively” with the White House or “appreciate the ideological differences.”

Suffice to say, if this dispute is solely about a security vulnerability, it is unclear how the “ideological differences” between the Trump administration and Anthropic’s liberal leadership would matter. Nevertheless, Axios goes on to report that Anthropic compounded its own difficulties by soliciting a review of Amazon’s research from Luta Security’s Moussouris, whom the administration views as a “radical Democrat.”

Again, if the export controls are motivated exclusively by cybersecurity concerns, then Moussouris’s ideological leanings would seem irrelevant. 

In context, it is hard not to read the administration’s complaints about Anthropic’s failure to “communicate” as demands for the company to genuflect before Trump.  

Process is important

All this said, Amazon’s research is not currently available for public scrutiny. We do not know exactly what Fable’s vulnerabilities are, nor precisely what administration officials were thinking when they effectively banned the model. 

What’s certain, however, is that the process behind the Fable ban was grievously flawed. The administration has not formulated any objective and binding standards for AI model safety — much less, gotten Congress to ratify such requirements. 

Nor did it conduct any thorough or transparent cost-benefit analysis before unilaterally removing Fable from the market, as regulatory agencies typically must before enacting sweeping policy change. And the potential costs of the Fable crackdown aren’t negligible: For example, if foreign businesses know that the US president can (and will) revoke their access to American AI models on a whim, then they will have an incentive to replace Claude and ChatGPT with non-American alternatives. 

Perhaps Amazon identified a liability serious enough to override such concerns. But the administration has made little effort to establish as much. 

We need a better alternative to the robot apocalypse

AI models are growing rapidly more powerful — and thus, more dangerous. It is possible that AI progress will have positive or neutral implications for cybersecurity: Advanced models could end up doing as much or more to shore up defenses as to undermine them. 

But that is not guaranteed. 

To mitigate the risks that frontier AI systems present, the government may be justified in establishing licensing processes that condition a new model’s release on its compliance with safety standards. 

There is a difference, however, between Congress establishing an impartial, rule-bound regulatory process and the executive branch banning AI systems at will. If tech CEOs shouldn’t have full discretion over which models get released, presidents must not have unchecked authority over which get blocked. The alternative to reckless, AI accelerationism should not be capricious cronyism — but, for now, it appears to be.