Fable5が強制的にオフラインになる前の24時間、ホワイトハウスで何が起きたのか？

Original title: Inside the whirlwind 24 hours that led the White House to slap export controls on Anthropic

Original authors: Sophia Cai and Cheyenne Haslett, Politico

Original translation/compilation: Peggy

Editor's note: The Fable 5 model, released by Anthropic just a few days ago, was suddenly taken offline. The trigger was the Trump administration citing national security concerns to restrict foreign nationals from accessing Anthropic's advanced AI models, Fable 5 and Mythos 5. Due to the difficulty in precisely distinguishing user identity and access permissions, Anthropic ultimately chose to disable the models for all customers.

Behind this sudden takedown lies a high-pressure negotiation between the White House and Anthropic that unfolded over 24 hours. Concerning whether the safety guardrails of Fable 5 could be bypassed, Anthropic CEO Dario Amodei engaged in several tense phone calls with senior government officials, including the Treasury Secretary, the Commerce Secretary, and the White House Cyber Director. The government believed the model could be used to identify software vulnerabilities, posing a national security risk; Anthropic, however, emphasized that the issue was not a 'general jailbreak' and criticized the government's actions for lacking a transparent, clear, and technically factual legal process.

The crux of this incident is not just a model takedown, nor is it merely a breakdown in communication between the White House and an AI company. Rather, it signifies that advanced AI models are being redefined as potential national security assets. When model capabilities enter sensitive areas like cybersecurity, vulnerability discovery, and intelligence analysis, the government is no longer satisfied with corporate self-assessment and voluntary review. Instead, it is beginning to employ stronger policy instruments, such as export controls, to directly intervene in the release pace of models.

Ironically, Anthropic has been one of the most vocal advocates for regulation within the AI industry, yet it has now become the target of the regulatory hammer. The government believes Anthropic did not treat the security vulnerabilities seriously enough, while Anthropic believes the White House overreacted. Behind this dispute lies a new question the AI industry must face in the future: Who has the right to determine if a model is safe enough? Is corporate self-certification of safety still sufficient? And how will rapid interventions justified by national security change the compliance costs, release schedules, and global competitive landscape for AI companies?

The Anthropic incident could become a watershed moment for advanced AI regulation. In the future, the release of cutting-edge models will no longer be just a product launch, but more akin to a high-risk deployment involving national security, geopolitics, capital markets, and industrial order.

Below is the original text:

Before the Trump administration imposed sweeping export controls on Anthropic, senior White House officials engaged in a round-the-clock, 24-hour mediation effort to persuade the AI company to voluntarily take down a newly released model, which multiple government officials believed posed a national security risk.

According to two government officials and one senior White House official, during this turmoil, Anthropic CEO Dario Amodei had several tense phone calls with high-ranking government officials, including U.S. Treasury Secretary Scott Bessent and White House Cyber Director Sean Cairncross. These individuals requested anonymity as they were not authorized to discuss the matter publicly.

This event also brings an increasingly pressing issue to the forefront: How should the White House draw the line between innovation and security when faced with rapidly iterating advanced AI models whose capabilities are constantly expanding?

The details of these phone calls have never been reported before.

The Night Before the Model Takedown: Security Dispute

Following the government's imposition of export controls, Anthropic was forced to take down its new model, Fable, which had been publicly available for only a few days. Anthropic had previously assured the public that the model was safe and controllable, but shortly after its release, senior government officials began reassessing whether its safety guardrails were as robust as the company claimed.

Two government officials and a senior White House official stated that on Thursday, two days after Fable's public release, Amazon CEO Andy Jassy expressed concerns to the White House, stating that the model's safety guardrails might be at risk of being bypassed.

A person familiar with Amazon's communications said the company was responding to a government request to provide feedback on potential risks.

By Friday morning, the issue had escalated to the highest levels of the White House.

According to the government officials and the senior White House official, Bessent, Cairncross, White House Chief of Staff Susie Wiles, and other senior officials held a meeting to discuss the model and the government's response. One of the sources said Bessent participated remotely as he was traveling to Houston for a scheduled public event.

After the meeting, the government attempted to contact Amodei but was told he was temporarily unavailable to take calls due to attending a health retreat. This was according to one government official and one senior White House official.

An Anthropic spokesperson denied this claim, stating, "That is completely untrue."

A person close to Anthropic said the government first requested to speak with Amodei around noon, and Amodei was on the phone with senior officials within 1 hour and 15 minutes. This person stated that during the time Amodei was initially unavailable, Anthropic offered to have other executives communicate on his behalf.

Once the government finally reached Amodei, he participated in three phone calls. The senior White House official and a government official said about half a dozen senior government officials participated in the calls, including Cairncross, Bessent, and Commerce Secretary Howard Lutnick.

The senior White House official added that other White House staff and government officials participated in some of the calls, including Commerce Under Secretary for Industry and Security Jeffrey Kessler, White House Cabinet Secretary Will Scharf, Deputy Chief of Staff Richard Walters, and Presidential Policy Advisor Walker Barrett.

During the calls, Amodei attempted to clarify what he believed were misunderstandings. He rebutted the government's security concerns, defended Anthropic's safeguards, and emphasized that the method used to bypass the protections was specific to certain scenarios and did not constitute a broader 'jailbreak' – meaning it did not completely free the model from the restrictions of Anthropic's safety guardrails.

In a blog post published after the export controls were implemented, Anthropic stated: "Currently, no tester has found a universal jailbreak method – one that could broadly circumvent the model's safety measures and unlock a wide range of cyber capabilities." The company also said that it is currently impractical for Anthropic, or any other AI company, to completely prevent all jailbreak attempts.

Anthropic also defended its own safety systems, claiming their protection mechanisms are "so strong that many users complain they are too broad."

However, Cairncross and Bessent were not convinced by Amodei's explanations. A White House official stated that Amazon's findings had been submitted for review by the National Security Agency, which considered the materials sufficient to constitute "evidence."

According to the senior White House official and two government officials, the government urged Anthropic to voluntarily take down the model and coordinate with the government to fix the vulnerability. Amodei requested more time and information but did not commit to taking down the model. The senior White House official said that during one of the calls, Bessent directly told Amodei that he was making "a bad decision."

The Regulatory Hammer Falls: Fable 5 and Mythos 5 Face Controls

Shortly after the calls concluded, the Trump administration imposed export controls on both the Fable 5 and Mythos 5 models. Anthropic stated that the government invoked national security authority to prohibit foreign nationals from using these models. The company said the "practical effect" of this order was that, to ensure compliance, Anthropic had to "suddenly disable" access for all customers to the affected models.

A senior White House official stated: "We begged them to work with us for hours. Export controls were the last resort. This is not something we wanted to do, but we felt we had no choice."

After this report was published, a person close to Anthropic countered, claiming the company was not given a choice for "voluntary cooperation."

This person stated: "The White House only gave us 90 minutes to take down the models, without providing any details about the actual threat. There was never any 'begging' or request for us to work with them. Just an announced 90-minute ultimatum."

A senior White House official said White House officials had heard Amodei compare the dangers of Anthropic's technology to nuclear bombs. Therefore, government officials found it difficult to understand when this CEO, who has consistently advocated for AI safety regulation, was reluctant to take down the system to fix known security vulnerabilities.

Anthropic has long been seen as one of the most vocal advocates for regulation within the AI industry. The company has repeatedly argued that a regulatory framework is necessary to address the potential global security risks and job displacement resulting from the rapid development of AI.

Three people familiar with the government's thinking said Amazon was not the only company that raised concerns with the government.

One of them stated: "The core issue is that Anthropic didn't take this seriously enough. If Anthropic had taken it seriously, instead of downplaying it as an isolated incident, and proactively fixed it or suspended access, this never would have escalated to this point."

Another person close to Anthropic countered, arguing that the so-called "jailbreak" did not mean the security systems of Fable 5 had failed, and pointed out that the company had communicated with the government before releasing Fable. This person said that during multiple rounds of communication, the government did not object to the release of Fable.

After the government implemented export controls, Anthropic stated in a blog post that it would comply with the government's directive, but it also believed the action was an overreaction.

Anthropic stated: "As we have said publicly before, we believe the government should have the power to prevent unsafe deployments. However, this should be based on a transparent, fair, clear, and technically factual legal process. This action does not meet these principles."

A White House official, speaking anonymously as permitted, said innovation remains the White House's "primary goal, but we must also prioritize security."

Amazon declined to disclose specific details of its communications with the government in a statement. An Amazon spokesperson said: "It is not unusual for the government to seek our input on potential security risks. When such situations occur, we do not disclose the details of those discussions."

Earlier this April, Anthropic announced that its latest powerful model, Mythos, would only be made available to a select group of technology and cybersecurity companies to help them test software vulnerabilities. Anthropic stated at the time that due to the model's powerful capabilities, it could cause serious consequences if it fell into the wrong hands, thus necessitating restrictions on its release.

The debut of Mythos led to a series of meetings between Amodei and senior White House officials. Both sides described these meetings as productive. Subsequently, multiple rounds of discussions took place regarding how to regulate advanced AI models, eventually leading to a recent executive order requiring companies to voluntarily submit advanced models for government review before widespread deployment.

Fable 5, publicly released this week, was described by Anthropic as a "Mythos-level model" but equipped with additional safety measures allowing it to be opened to general users. The model had been reviewed by the U.S. government and the UK AI Safety Institute.

However, after the alleged security flaw was disclosed, multiple government officials believed the model had to be taken down immediately.

On Saturday morning, former White House AI director and staunch anti-regulation advocate David Sacks posted on X supporting the government's decision to impose export controls on Anthropic.

Sacks stated that he did not view this "jailbreak" as simple or insignificant, nor did he believe the export controls were an attempt by the government to broadly control the AI industry. Sacks wrote: "The government now wants Anthropic to fix the security issue, then lift the export controls, and allow Fable to return to public release. The government wants this to happen as soon as possible. Frankly, the government is puzzled that Anthropic, which has said security requests are its top priority, is now unwilling to cooperate."

Sacks and other officials within the government have previously criticized Anthropic multiple times, accusing it of having left-wing political bias and of fear-mongering by advocating for stronger industry regulation and warning of massive job displacement.

Earlier this year, Defense Secretary Pete Hegseth and the Pentagon escalated the disagreement between the government and Anthropic to an unprecedented level. On March 3, because Anthropic refused to allow its AI tools to be used for mass domestic surveillance and autonomous weapon systems, the Pentagon listed the company as a supply chain risk.

On Saturday, Sacks stated that the pre-existing dispute between the government and Anthropic was separate from the export control decision. Sacks wrote: "The government values Anthropic's technical capabilities and believes that while this issue is serious, it should be easily resolved. The ball is now in Anthropic's court."

Original Link