Amazon’s Andy Jassy Triggered Washington’s Crackdown on Anthropic’s AI

Amazon Chief Executive Andy Jassy picked up the phone to Treasury Secretary Scott Bessent on June 11, and within a day the United States government forced Anthropic to shut down two of its most advanced AI models worldwide, the first time Washington has used export law to pull a commercial AI product offline. The models, Claude Fable 5 and Claude Mythos 5, went dark for every user on the planet, including Anthropic’s own foreign staff.

The twist that got less attention than the jailbreak itself: Amazon is Anthropic’s largest financial backer, has poured roughly $13 billion into the startup, and runs the cloud infrastructure Claude depends on. Its chief executive’s phone call is what set the shutdown in motion.

The Call That Started It

Anthropic released Fable 5 and its more powerful sibling Mythos 5 on June 9, its first public versions of a model family the company had previously said was too dangerous in cybersecurity terms to hand to the world.

Amazon’s own researchers tested it almost immediately. They found a series of prompts that got Fable 5 to hand over information useful for cyberattacks, the kind of output the model was supposed to refuse. Jassy learned of the finding and, during a previously scheduled call with White House officials on Thursday, June 11, mentioned it. Officials told him to raise it directly with Bessent, and he did, that same day.

Bessent has led the administration’s response to Anthropic’s Mythos-class models, largely over fears that Mythos-powered cyberattacks could threaten the global financial system. His call with Jassy triggered a chain of events that became international news within 48 hours.

By Friday evening, Commerce Secretary Howard Lutnick had sent Anthropic co-founder and CEO Dario Amodei a letter demanding the company fix the flaw or pull the models, giving the company roughly 90 minutes to respond. Anthropic disabled both by 10 p.m. Eastern time.

We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.

Anthropic wrote that in a blog post hours after the shutdown, and confirmed the news in a statement disabling both models for every customer, adding that access to every other Claude model was untouched.

Why Would Anthropic’s Biggest Investor Turn Informant?

Amazon is not a bystander in Anthropic’s story. It has committed roughly $13 billion to the company since 2023 and Anthropic has separately pledged more than $100 billion in spending on Amazon Web Services infrastructure, including Trainium chip capacity. Claude runs on AWS Bedrock. Amazon profits every time a customer uses it.

That is what made Jassy’s call so awkward inside Amazon itself. One Amazon engineer, speaking anonymously, joked that colleagues were calling it out for having “snitched” on Anthropic, and others said they never got the chance to try Fable 5 before it vanished.

Amazon has offered a carefully worded, consistent line on the episode. An Amazon spokesperson said governments routinely ask cloud providers for security input and that the company keeps those conversations confidential.

Jassy did not appear to be trying to sink Anthropic’s product. He told Bessent he was worried about cyber capability across every frontier lab, not just Anthropic’s, according to people familiar with the call. But once the finding reached the Treasury Department, the outcome was out of Amazon’s hands. Whether Jassy expected a global shutdown is unclear; what is clear is that Anthropic’s own largest investor supplied the evidence used to take its product off the market.

What Did the Export Order Actually Prohibit?

The Commerce Department’s directive barred any foreign national, anywhere, from accessing Fable 5 or Mythos 5. That included foreign nationals physically inside the United States and Anthropic’s own non-citizen employees. The government relied on a “deemed export” standard, which treats handing an AI model to a foreign national as equivalent to shipping the technology overseas, regardless of where that person is sitting.

Anthropic had no way to check a user’s citizenship in real time. With no technical means to separate foreign accounts from domestic ones, the company said it had one option: take both models offline for everybody, everywhere, until it could comply.

The order named national security authorities but gave few specifics. Anthropic said it understood the concern to be the jailbreak Amazon had found, though the government never confirmed that directly to the company.

Where Amodei and the Administration Disagree

Dario Amodei argued on calls with senior officials that the flaw Amazon found was narrow, capable of unlocking cybersecurity output in one specific way rather than defeating Fable 5’s safeguards broadly. David Sacks, a former White House AI adviser who now co-chairs a presidential science and technology council, gave a different account on social media, saying a trusted partner of both Anthropic and the government had found a working jailbreak and that Amodei had refused to fix it or pull the model.

Anthropic later said its own testing showed the same technique worked against other companies’ models too, including OpenAI’s GPT-5.5 and the Chinese model Kimi K2.7, neither of which faced export restrictions. Katie Moussouris, founder of the security firm Luta Security, reviewed the technique at Anthropic’s request and concluded it described normal defensive security work rather than a true guardrail bypass.

• Anthropic and Dario Amodei say the flaw exposed only a handful of previously known, minor vulnerabilities, reproducible on rival models.
• David Sacks and the administration say the bypass enabled a working cyber weapon and that Anthropic dragged its feet on a reasonable safety fix.
• Alex Stamos, Facebook’s former chief security officer, reviewed the underlying research and wrote that the findings were “valid” but that nothing in them justified a reaction on this scale.

Bessent, for his part, told Amodei directly on a call that he was “making a bad decision” by resisting the government’s demands, according to Politico’s reporting on the standoff.

Sovereign AI Anxiety Spreads From Paris to Ottawa

The shutdown landed the same week G7 leaders met in Evian, France, and it became a talking point among heads of state. French President Emmanuel Macron called the episode a wake-up call about AI’s dangers, though he separately said the restrictions themselves amounted to a “bad thing” and were, in his words, “strictly nationalist.”

• Emmanuel Macron, French President, said the reaction was “strictly nationalist” even while calling it a warning worth heeding.
• Mark Carney, Canadian Prime Minister, said the episode showed the risk of over-reliance and that “nobody has done anything wrong,” but that allies still need to diversify.
• Henna Virkkunen, the European Union’s executive vice president for tech sovereignty, called the shutdown “another wake-up call” for Europe.
• Bruno Retailleau, a former French minister running for president in 2027, argued Europe must treat AI like nuclear power, writing that nations must “master it or suffer it.”

The episode fed a broader push already underway: some 46 countries are now funding roughly 135 sovereign AI projects of their own, a trend Beijing has moved on too, tightening its own grip on which AI models its own companies can send abroad in a mirror-image push detailed in Beijing’s own effort to wall off its top models. Critics of the U.S. order, including AI policy expert Dean Ball, who briefly served in the Trump administration, called the move “simply cartoonish” and said he could not tell whether it reflected genuine security concern or something closer to score-settling.

How Fable 5 Came Back Online

The standoff did not end quickly. Anthropic sent senior technical staff to Washington the weekend after the shutdown. On June 26, the government approved a small list of American cybersecurity firms to use Mythos 5 again. Full resolution took until June 30, when Lutnick announced the Bureau of Industry and Security had reviewed the diversion risks and would withdraw the controls entirely. Fable 5 returned to global users on July 1, the same day Anthropic launched a separate new model, Claude Sonnet 5.

As part of the deal, Anthropic said it trained a new classifier that blocks the specific jailbreak technique in more than 99% of cases, and government researchers from the Commerce Department’s AI standards office tested and endorsed the fix. Anthropic also agreed to build a shared framework with Amazon, Microsoft and Google for judging when a jailbreak is serious enough to demand action, a sign the crisis reshuffled how frontier labs and their own investors now police each other. The three-week gap also handed a competitive opening to rivals; models tied to firms in China moved to fill the space, an opportunity examined in this site’s look at how Chinese open-source labs gained ground during the blackout.

Anthropic’s IPO Clock Keeps Ticking

The shutdown landed at an inconvenient moment. Anthropic confidentially filed for an initial public offering in June, not long after disclosing a Series H round that valued the company near $965 billion and a revenue run rate approaching $47 billion.

• $13 billion is roughly what Amazon has invested in Anthropic since 2023, making it the company’s largest single backer.
• $100 billion is what Anthropic has committed to spend on Amazon Web Services infrastructure, including Trainium chips.
• $965 billion is Anthropic’s valuation after its Series H round, disclosed alongside its confidential IPO filing.
• $47 billion is Anthropic’s annualized revenue run rate as of May, a figure that has climbed sharply over the past year.

OpenAI, Anthropic’s closest rival, faced its own version of the same pressure. It agreed to limit early access to its GPT-5.6 model to a small set of government-vetted partners, a rollout detailed alongside Anthropic’s own enterprise push in OpenAI’s latest product rollout against Anthropic. OpenAI CEO Sam Altman said afterward, “This isn’t quite the process that we think is optimal.”

ThreatLocker Chief Executive Danny Jenkins argued the whole episode proved the limits of trying to legislate capability out of existence, saying stronger AI models “aren’t a genie that can be crammed back into a bottle.” Enterprises that build around any single AI vendor are drawing their own lesson, and some are already looking at cheaper in-house alternatives, a shift visible in Microsoft’s own push to cut reliance on outside labs with its newly unveiled in-house model lineup.

Anthropic is also still fighting the Pentagon in court over a “supply chain risk” designation handed down after the company refused to let its models be used for autonomous weapons decisions, a fight a federal judge partly paused with an injunction in March, now under appeal. Amodei has become enough of a political target inside the administration, largely over his AI safety advocacy and his 2024 support for Kamala Harris, that when Lutnick finally signed the letter lifting the export ban, it went to Anthropic co-founder Tom Brown, not to Amodei.

Frequently Asked Questions

Is Claude Fable 5 available again?

Yes. Anthropic restored global access on July 1, and gave Pro, Max, Team and select Enterprise subscribers up to 50% of their weekly usage limit on Fable 5 through July 7 before shifting the model to standard usage credits.

Why did the shutdown affect people outside the United States who had nothing to do with the jailbreak?

Because the export order used a “deemed export” standard covering any foreign national anywhere, and Anthropic’s platform had no real-time way to verify a user’s citizenship, so it could not selectively block only the accounts in question.

Did Amazon face any consequences for triggering the shutdown?

No formal consequences have surfaced. Amazon has stuck to a single public line, that governments routinely consult cloud providers on security risks and that those conversations stay private, even as some of its own engineers joked internally about the company having tipped off regulators on its own investment.

Is Anthropic still in a legal fight with the Pentagon?

Yes. That separate dispute, over a “supply chain risk” label tied to Anthropic’s refusal to loosen restrictions on autonomous weapons use, is still working through appeal after a federal judge granted Anthropic a preliminary injunction in March.

What changes going forward so this does not happen again?

Anthropic says it is building a shared framework with Amazon, Microsoft and Google for judging how serious a jailbreak needs to be before it triggers action, along with pre-release government access to future models and safeguards.