Days after Anthropic’s Dario Amodei argued that dangerous AI releases should be blocked, the White House imposed export controls on Anthropic’s latest model, forcing the company to take it offline. The move rests on a claimed security vulnerability tied to all
Dario Amodei wanted strict limits on cutting-edge AI—so badly that he argued in an essay last week that releases should be “blocked or reversed as a threat to public safety” if they fail to meet strict security standards.
Then, days after his manifesto went live, Anthropic’s latest model went dark.
The administration’s action targeted the model in a way that effectively pulled it from public reach. In its unrestricted form. the system is known as “Mythos.” In its heavily bounded. publicly accessible version. it is called “Fable.” The company had previously shared Mythos exclusively with vetted public and private organizations so they could “steel their cyber-defenses” against its capabilities. Before releasing the general-public version. Anthropic lined Fable with strict safety guardrails—among other restrictions. it would refuse to answer virtually any query about cybersecurity or biology. aimed at preventing its use for hacking and bioterrorism.
But after the White House learned of what it described as a potential security vulnerability, it imposed export controls on Friday. The practical effect was immediate and sweeping: it became unlawful for Anthropic to provide Fable to any foreign national, including its own immigrant employees.
Because AI models still can’t scan a user’s brain and confirm their nationalities, the administration’s order meant Anthropic had to take Fable offline completely.
Amodei is not celebrating. And the broader community that focuses on “AI safety” is uneasy for a reason that goes beyond this single model: proponents of tighter guardrails say the country needs predictable rules that mitigate risks while allowing benefits—not a regime shaped by executive discretion.
The administration’s case for banning Fable is anchored in the idea that it could pose unique security challenges, even with guardrails. Anthropic had itself been unnerved by Mythos’s apparent gifts for cybercrime, and Fable—its bounded version—remained exceptionally powerful.
The trigger for the crackdown began earlier. after a warning from one of Anthropic’s own investors that Fable was vulnerable to a potential “jailbreak. ” a method for circumventing the model’s safety controls. Last Thursday, Amazon—holding a $13 billion stake in Anthropic—shared research with administration officials documenting such a jailbreak.
The White House responded by reaching out to Anthropic and asking it to fix the issue. Anthropic, in turn, insisted its model was safe and said the administration was misunderstanding Amazon’s research.
From there, the administration concluded that Anthropic was unable or unwilling to fix the problem. It then determined that export controls were the only way to ensure Fable would not degrade America’s cybersecurity.
Yet critics say this story is incomplete—and that the administration’s response may be selective. One of the key disputes centers on whether Fable’s alleged vulnerability is truly unique.
Katie Moussouris. head of the cyber security group Luta Security. reviewed a copy of Amazon’s findings and said they raised no novel risks. In her description. when Fable is prompted in a certain way. it would identify software vulnerabilities—ostensibly to help a user shore up their defenses. Moussouris argued that many frontier models, including OpenAI’s GPT 5.5, will provide the same service.
Anthropic disputes the framing. The company says it subjected Fable to thousands of hours of testing—by independent organizations and by the US government—to ensure the model contains no universal jailbreak. defined by Anthropic as “a method that can very broadly bypass the model’s safeguards. unblocking a wide range of cyber capabilities.” Anthropic also argues that narrow jailbreaks like the ones Amazon described are impossible to fully preempt.
If Anthropic is right, then critics say the administration’s targeting of Fable would look less like a hard security necessity and more like something arbitrary.
There’s also a second thread running through the criticism—one tied to previous friction between Anthropic and the Trump administration. Earlier this year. Anthropic and President Donald Trump’s Defense Department collided after Anthropic refused to approve the use of its models for mass surveillance and fully autonomous weapons systems. The Pentagon then declared Anthropic a “supply chain risk. ” a designation intended to restrict the capacity of government contractors to do business with it.
The policy was described as legally dubious and transparently disingenuous in the account being reported: the administration was said to be asserting that Anthropic’s AI was structurally unsafe for government work while continuing to use that same AI for government work. The stated intention, critics say, was to punish Anthropic for insisting on contractual terms the administration did not like.
For critics of the export controls, that history is a warning sign about impartiality. They also point to the White House’s apparent comfort with two of Anthropic’s top competitors—OpenAI and Elon Musk’s xAI—as additional cause for skepticism.
The strongest evidence offered for bad faith, in this account, comes from the administration’s own explanations. In an interview with Axios. a “source familiar with the administration’s thinking” said Anthropic’s difficulties partly reflected its inability to “communicate effectively” with the White House or to “appreciate the ideological differences.”.
If the issue were purely about a security vulnerability, the criticism argues, “ideological differences” would not matter. Still. Axios also reported that Anthropic compounded its difficulties by soliciting a review of Amazon’s research from Luta Security’s Moussouris. whom the administration views as a “radical Democrat.” Again. the critique says that ideological leanings shouldn’t be relevant if the export controls are driven solely by cybersecurity concerns.
To be clear, Amazon’s research is not publicly available. The vulnerabilities described by Amazon—and exactly what administration officials were thinking when they effectively banned the model—remain unclear.
But what is clearer. critics say. is the process: the administration did not formulate objective and binding standards for AI model safety. and it did not get Congress to ratify such requirements. The account argues that the administration also did not conduct the kind of thorough and transparent cost-benefit analysis regulators typically perform before sweeping policy change.
It also argues that the costs could be meaningful. If foreign businesses come to believe that the US president can revoke access to American AI models on a whim, they may choose to replace Claude and ChatGPT with non-American alternatives.
The argument then widens to what comes next, and it centers on who should hold the power. AI models are moving quickly toward more dangerous capabilities, and while advanced systems might someday strengthen cybersecurity, that outcome isn’t guaranteed.
In the view presented here. the government may have justification for licensing processes that condition a new model’s release on compliance with safety standards. The crucial distinction, critics say, is between Congress creating a rule-bound process and the executive branch banning systems at will.
If tech CEOs shouldn’t have full discretion over which models get released, the criticism goes, presidents should not have unchecked authority over which models get blocked.
For now, Fable is offline, export controls have reshaped access, and a debate that once lived mainly in essays and security briefings has landed in the real world—fast enough that the timing itself feels like a warning.
Anthropic Dario Amodei Fable Mythos AI regulation export controls White House cybersecurity jailbreaking Katie Moussouris Luta Security Amazon investment