Amazon Poured $13 Billion Into Anthropic, Then Helped Get It Banned Worldwide

Source: Patrick Boyle | Published: 2026-06-20T11:30:06Z

The so-called jailbreak flagged as a national security threat turned out to be asking an AI coding assistant to read code and fix bugs — literally the first feature advertised on Anthropic's own product page.

This was less than two weeks after Anthropic had confidentially filed its IPO paperwork with the SEC. The prospectus disclosed annualized revenue of $47 billion and an expected valuation of $965 billion.

What You Can Do in 90 Minutes

When the notice arrived, Anthropic's engineers faced a problem with no clean solution: accurately identify and block every "foreign national" among hundreds of millions of global users — in 90 minutes.

Export control law has a mechanism called "deemed export" — an engineer on a foreign visa logging into a work account at the San Francisco office is legally equivalent to exporting that technology to their home country. This meant the non-U.S. engineers who had spent the past year building these models suddenly found themselves locked out of their own systems by their own company.

No technology exists to perform real-time nationality verification at that scale in 90 minutes. Anthropic did the only thing that could achieve compliance: shut it down for everyone. Fable 5 and Mythos 5 went offline globally, simultaneously.

That "Highly Dangerous Jailbreak"

White House AI adviser David Sacks explained the shutdown on social media: a "trusted partner" had discovered a highly dangerous jailbreak technique capable of bypassing Fable 5's safety guardrails. He claimed the government had asked Anthropic to fix the vulnerability, and Anthropic refused.

Anthropic then released its own account. From their perspective, the letter arrived at 5:21 PM with no technical details and no record of any conversation in which they had refused to implement a fix. Their statement quoted their own assessment of the supposed "threat":

"The technique essentially involves asking the model to read a specific codebase and fix software bugs within it."

So the jailbreak classified as a national security threat was: asking an AI coding assistant to read code and patch bugs. That's the first feature listed on Anthropic's product page.

The irony runs deeper. The security community's actual assessment of Fable 5 was the polar opposite. Prominent security researcher Valentina Palmiotti reported that the model became a community joke on launch day for being excessively cautious — it would refuse to read an ordinary IT blog because the post contained words that sounded "vaguely related to computer networks." Matt Suiche, a member of the technical staff at AI cybersecurity startup Tolmo, told TechCrunch that simply asking Fable 5 to write secure code caused it to automatically downgrade to a weaker model, because it determined you were engaged in "dangerous cybersecurity work."

The reality: this company spent thousands of hours on red-teaming, trained its product to be so cautious it bordered on useless, and still got shut down by the government in 90 minutes.

The Person Who Made the Call

The "trusted partner" Sacks referred to was Amazon CEO Andy Jassy.

A few numbers help put this in context. Amazon has already invested $13 billion in Anthropic, with commitments of up to $25 billion more. Meanwhile, Anthropic is one of Amazon's largest cloud customers — CNN reported in April that Anthropic had committed to paying AWS over $100 billion in service fees to train and run its AI models.

So the full logic of what happened is this: Anthropic raised billions from a cloud computing giant, signed a contract promising to spend $100 billion renting their servers, and then that giant's CEO called the White House on a Friday afternoon to ensure Anthropic's flagship product disappeared from the market.

Amazon also develops its own competing AI models.

Amazon's official statement read: "It's not unusual for governments to consult us on potential security risks. When that happens, we don't share the details of those discussions."

The Essay Dario Published the Day Before

The timeline contains a stranger coincidence.

On June 10th — 24 hours after Fable 5 launched and less than 24 hours before the Commerce Department letter arrived — Anthropic CEO Dario Amodei published a detailed policy essay. In it, he compared the intersection of artificial intelligence and political institutions to a scene from The Lord of the Rings: hobbits trying to rouse Treebeard to defend the forest, but "Treebeard moves much more slowly than the hobbits." His argument: governments need transparent, fair, technically grounded statutory procedures for halting unsafe model deployments when necessary.

He called for regulation. The next day, it arrived.

This isn't meant as mockery. It illustrates a very specific trap: when you spend years telling regulators you've built a technology capable of threatening civilization, they don't sit down to contemplate the philosophical implications of recursive self-improvement. They give you 90 minutes to shut it off.

The Air at the G7 Lunch

Days later, the G7 summit convened in France. Organizers arranged a working lunch with Dario Amodei, OpenAI's Sam Altman, and Google DeepMind's Demis Hassabis. The topic: AI's opportunities for the financial industry.

Before the silverware was set, Goldman Sachs had already cut its Hong Kong employees' access to Anthropic products. JPMorgan followed shortly after. Both banks' legal teams glanced at Anthropic's terms of service — which technically excluded usage in Greater China — and decided it was safer to quietly remove Claude from the employee tools dropdown before facing potential compliance exposure on Monday morning.

This is the real commercial damage from the episode: not technical, but reputational. Financial institutions had spent months negotiating global usage contracts. One letter made those contracts legally precarious overnight. French President Macron said after the summit that the incident "clarified the G7's real situation" — if the United States can flip the switch on any given day, what does that mean for the multi-trillion-dollar enterprises that depend on American AI services?

I didn't expect to see a French leader criticizing American commercial overreach. But here we are.

5:21 PM

Anthropic's shutdown timestamp was 5:21 PM.

The following day, a Chinese AI lab called Zhipu announced a new open-source model release. They scheduled the announcement for 5:21 PM.

The symbolism is pointed. Chinese AI companies offer open-source models that can be self-hosted — meaning they are structurally immune to being arbitrarily switched off by Howard Lutnick on a Friday afternoon. DeepSeek's flagship model is currently priced at $0.87 per million output tokens, roughly one-sixtieth the cost of Anthropic's Fable 5.

According to Financial Times data, in the first two weeks of June, Chinese models processed twice as many tokens as their American counterparts among the top 20 models on OpenRouter; four of the top five came from China.

A Valuation Built on One Assumption

Anthropic and OpenAI's near-trillion-dollar valuations rest on a single core thesis: AI is a winner-take-all market, and the first movers will capture nearly all the profits the way Google captured search and Facebook captured social. Investors are betting that once LLM markets consolidate into a monopoly, these companies can charge the global economy a premium for access.

But if depending on American frontier models means a regulator can pull the plug on a Friday afternoon over a line of code, the logic breaks. The addressable market shrinks from "the global economy with internet access" to "people who clear background checks in Washington." The $965 billion valuation needs to be recalculated.

And if open-source models are already good enough, sixty times cheaper, and can't be shut down by anyone — then the productivity gains from the AI boom flow to the businesses using the models, not to the investors who spent billions building them and bet on a monopoly.

Anthropic spent years explaining to governments exactly how dangerous their products were, trained those products to be so cautious they wouldn't even read a blog post, and then was genuinely surprised when the government believed them.