一紙禁令，三通電話，一個失控的敘事：Anthropic的至暗24小時

On June 13, the U.S. government pressed pause on two of Anthropic's most advanced models.

One is Fable 5, the other is Mythos 5. The former was just publicly released, while the latter is targeted at more restricted cybersecurity clients. The ban came from the U.S. Department of Commerce, covering clients outside the U.S. as well as foreign nationals within the U.S. Ultimately, Anthropic's simple choice was to take them all offline.

Looking at all the details surrounding this incident, we have roughly pieced together the timeline of these 24 hours.

On Thursday, June 11, two days after Fable 5's public release, Amazon CEO Andy Jassy raised a risk to the White House. He was concerned that Fable 5's safety guardrails could be bypassed. Amazon researchers allegedly used a series of prompts to get Fable 5 to output information that should have been restricted, information that could be used for cyberattacks.

By Friday morning, June 12, the issue had reached the highest-level meetings at the White House. Treasury Secretary Scott Bessent, White House Cyber Director Sean Cairncross, White House Chief of Staff Susie Wiles, and other senior officials participated in the discussions. Bessent was en route to Houston at the time and joined the meeting remotely.

Then came three phone calls.

When Anthropic CEO Dario Amodei entered the call, he was facing about half a dozen senior officials. In addition to Bessent and Cairncross, Commerce Secretary Howard Lutnick was also present. Other participants included Deputy Commerce Secretary for Industry and Security Jeffrey Kessler, White House Staff Secretary Will Scharf, Deputy Chief of Staff Richard Walters, and Deputy Assistant to the President for Policy Walker Barrett.

Amodei tried to explain the situation as a misunderstanding. He believed that what Amazon had discovered was a specific bypass method, not a universal prompt jailbreak that could broadly dismantle safety guardrails. Anthropic later publicly stated that testers had not yet found a way to broadly bypass the model's security system.

However, the White House was not convinced.

The findings from the Amazon CEO were sent to the U.S. National Security Agency for assessment, and the White House believed it had enough evidence. The government requested that Anthropic voluntarily take the models offline and work with the government to patch the vulnerabilities. Amodei wanted more time and information but did not commit to removing the models. Bessent said directly on the call that he had made a "wrong decision."

Subsequently, export controls were imposed.

Anthropic presented a different narrative. They said the White House gave them only 90 minutes to take the models offline without explaining the specific threat details. The White House, on the other hand, stated that export controls were a last resort after hours of unsuccessful attempts to get Anthropic to cooperate.

Another key point in this matter is Amazon's delicate position in all of this.

At the end of 2024, Amazon made an additional $4 billion investment in Anthropic, bringing the total investment to $8 billion. Anthropic also designated AWS as its primary training partner, and future model training and deployment will utilize AWS chips. Claude has consistently been one of the most important models on Amazon Bedrock.

The alliance between Microsoft and OpenAI is already clear, and Amazon's bet on Anthropic was originally a way to circumvent that.

Microsoft has OpenAI. Google has Gemini and also invests in Anthropic. Amazon lacks a sufficiently strong in-house frontier model and can only tether the computing power of AWS, Trainium chips, and the Bedrock platform to an external model company.

But a year and a half later, Amazon and OpenAI are also connected.

This year, Amazon discussed investing up to $50 billion in OpenAI. OpenAI was seeking up to $100 billion in new funding at the time, and a potential deal could include OpenAI purchasing Amazon's AI chips. Axios also mentioned that OpenAI's annualized revenue for 2025 exceeded $20 billion, but its expenditure commitments reached $1.4 trillion.

Amazon needs frontier model companies to consume AWS computing power, validate its self-developed chips, fill its data centers, and place the strongest models on its enterprise cloud shelves. This is no longer a pure financial investment.

Thus, it both invests in Anthropic and draws closer to OpenAI. It acts as both a financier and a supplier to model companies. It must help them sell their models and explain to the government how dangerous these models are.

In terms of outcome, this time Amazon stood opposite Anthropic. From Anthropic's perspective, a partner providing money, cloud services, chips, and distribution channels handed over a security signal to the government sufficient to trigger a ban. Of course, Amazon's own account is, "The White House asked me; I was just responding to their questions."

Over the past two years, AI companies have liked to package themselves as national assets. The stronger the capability, the higher the valuation, the smoother the fundraising, the more imaginative the government procurement. Anthropic is particularly adept at this narrative. It distinguishes itself from OpenAI with more cautious security language and proves to regulators with the rhetoric of "frontier risk" that it should be taken seriously.

Now, the U.S. government truly treats models as national security assets.

The White House officials' bewilderment stems from here as well. Politico reported that officials on the White House side had heard Amodei compare the danger of Anthropic's technology to a nuclear bomb. When he refused to take the model offline due to a known security vulnerability, government officials saw it not as a technical disagreement but an issue of attitude.

This is not the first conflict between the two sides. On March 3, the Pentagon listed Anthropic as a supply chain risk because Anthropic refused to allow its AI tools to be used for large-scale domestic surveillance and autonomous weapons.

The conflict between Anthropic and U.S. officials has existed before.

This time, Anthropic stated that the government's directive did not specify the specific national security concerns and criticized the operation for lacking transparency, clarity, and a legal process based on technical facts. Anthropic believes the issue at hand is more akin to a narrowly scoped bypass method, insufficient to warrant such a broad ban.

However, from the government's perspective, model security is no longer an internal process where companies write white papers, conduct their own red teaming, and release their own system cards. Who can access the models, who can train them, and whether foreign employees can see the model weights are all being pulled into the language of export controls.

When Anthropic announced in April that Mythos would only be opened to limited technology and cybersecurity companies, it had already held multiple rounds of meetings with the White House. Before Fable 5 went live, it was also reviewed by the U.S. government and the UK AI Safety Institute. Anthropic claims the government did not raise objections before the model's release.

This made the conflict even uglier.

Before model release, it was about safety cooperation. After model release, it's about national security.

OpenAI is watching this incident from the sidelines.

With Anthropic forced to take its strongest models offline, OpenAI's relative position becomes more comfortable. The more Anthropic is entangled by regulatory issues, the easier it is for OpenAI to become the "cooperative" option. If Amazon continues to move closer to OpenAI, it also adds another layer of hedging.

Of course, there is no public evidence indicating that Amazon acted to help OpenAI undermine Anthropic.

The sharper truth is that when frontier models enter a cycle of trillion-dollar capital expenditure, partnerships are inherently no longer clean. Cloud vendors invest in model companies, model companies purchase cloud computing power, governments inquire with cloud vendors about security risks, and competitors red-team each other within the same regulatory arena.

The roles of financier, supplier, distributor, and reviewer are increasingly being played by the same companies.

This is more significant than a specific prompt jailbreak.

On the night Fable 5 and Mythos 5 were shut down, Anthropic lost more than just access points to two models. It lost a degree of control over its own narrative.

Amazon's hand still rests on the AWS console. OpenAI's fundraising table hasn't dispersed. The U.S. government has already taken a seat in the front row of model announcements.