Amazon CEO Andy Jassy went to senior Trump administration officials this week and raised security concerns about Anthropic’s new AI models. Those conversations directly set in motion the export controls that forced Anthropic to kill Fable 5 and Mythos 5.
The man who runs the company that invested billions in Anthropic, that hosts their models on AWS, that builds chips for them, quietly picked up the phone and called Treasury Secretary Scott Bessent.
Here is what happened, why it matters, and why I cannot stop laughing at Anthropic right now.
The New Revelation
Stephanie Palazzolo at Axios broke the story on X this afternoon. The Wall Street Journal and The Information both confirmed the reporting with their own sources.
Here is what we now know:
- Andy Jassy raised concerns to senior Trump officials this week
- He specifically discussed the issue with Treasury Secretary Scott Bessent
- Amazon’s own researchers found a way to jailbreak Fable 5 using a “series of prompts”
- Those conversations set in motion the export controls that killed both models
- The Commerce Department sent its directive to Anthropic on June 12 at 5:21pm Eastern
The government acted because the CEO of Amazon told them to.
How We Got Here
I covered the initial story yesterday. The export controls. The scope. Anthropic’s response. What it means for users. That was the what.
This is the who.
The chain of events goes like this:
Anthropic launched Fable 5 and Mythos 5 on June 9. Within 48 hours, a public researcher known as Pliny the Liberator claimed a jailbreak. Multi-agent decomposition. Unicode tricks. Narrative framing. The model produced step-by-step stack buffer overflow exploits for x86 Linux.
Amazon’s researchers found a similar bypass around the same time. Except they did not post it on X. They escalated it internally to their CEO.
Andy Jassy took it directly to Washington.
He talked to Treasury Secretary Scott Bessent. He may have talked to Commerce Secretary Howard Lutnick too. He told them this was a national security risk.
On June 12 at 5:21pm Eastern, Lutnick sent a letter to Anthropic CEO Dario Amodei. The directive: suspend all access to Fable 5 and Mythos 5 for every foreign national on the planet, including Anthropic’s own non-US employees.
Anthropic could not filter by citizenship in real time. So they shut both models down for everyone.
From launch to death: 72 hours.
The Betrayal Angle
This is the part that kills me.
Amazon has invested billions of dollars in Anthropic. Initial $4B. Another $8B. Plans to go north of $25B total. AWS is Anthropic’s primary cloud provider. Anthropic uses Amazon’s custom Trainium chips for training. Claude is available through Amazon Bedrock. Amazon has a board seat.
And the CEO of Amazon still went to the White House and said “this thing is too dangerous to exist.”
It is hard to overstate how wild this is. Imagine if Microsoft’s CEO called the government about OpenAI’s latest model. Or if Google’s CEO took concerns about DeepMind directly to the Treasury Secretary. The markets would lose their minds.
But here we are. Anthropic’s biggest strategic partner and largest investor personally trigger the kill switch on their best product.
What Amazon Actually Found
The jailbreak Amazon identified was not exactly a civilization-ending threat.
According to reporting from the WSJ, Amazon researchers used “a series of prompts” to get Fable 5 to produce information that could be used in cyberattacks. Specifically, they got the model to identify a small number of software vulnerabilities.
Anthropic’s official response: those vulnerabilities were “previously known” and “minor.” They noted that other publicly available models “can discover them as well without requiring a bypass.”
Katie Moussouris, CEO of Luta Security, reviewed the evidence and called the government’s response a “complete overreaction.”
Here is the part that makes this truly funny. Anthropic claimed over 1,000 hours of pre-launch red-teaming found “no universal jailbreak.” They built safety classifiers that silently hand off risky queries to the weaker Claude Opus 4.8. They restricted Mythos 5 to “vetted partners” through Project Glasswing.
They did everything they could to be the “safe” AI company.
And their own investor’s security team found a bypass anyway. And their own investor’s CEO used that bypass to get the government to shut them down.
The Self-Inflicted Drama (aka My Favorite Part)
Anthropic spent years positioning as the safety-first AI company. They refused to open-source their models. They belted Fable 5 with safety classifiers. They locked Mythos 5 behind a vetting process and told everyone it was too dangerous for the public.
They did all of that and it still was not enough.
Their own investor went behind their back and killed the product anyway. Not a competitor. Not a foreign government. The company that literally runs their cloud infrastructure.
You can not write this stuff.
If Anthropic had just shipped Fable 5 without the safety theater, without the “look how responsible we are” posture, without positioning Mythos as forbidden fruit, would Jassy have had a case to make? Hard to say. But the contrast is the point.
Anthropic built the safest cage in AI. And their biggest investor personally handed the government the keys.
OpenAI must be watching this and struggling not to laugh out loud. They do not have a single investor with a board seat who would go to the government and call their own model dangerous. Because that arrangement would be insane.
What This Means Going Forward
The Amazon-Anthropic relationship just took a shot that might be fatal.
Can Anthropic trust Amazon after this? Jassy did not warn them. He did not raise the issue privately first. He went straight to the Treasury Secretary and let the government land the punch.
Does Amazon keep its board seat? Does Anthropic start moving workloads off AWS? Anthropic was reportedly planning an IPO. Killing your flagship model a week before potential IPO paperwork is catastrophic timing.
The broader question for the industry is uglier. Any major investor can now trigger a government kill switch on an AI model. If Microsoft had a board seat at OpenAI, would a different administration have responded differently? Every frontier lab needs to ask themselves: who has access to the government that I do not control?
Bottom Line
Yesterday I told you the what. Here is the who.
Andy Jassy went to Scott Bessent. Scott Bessent talked to Howard Lutnick. Howard Lutnick sent a letter to Dario Amodei. And Anthropic’s best models died.
The most ironic part? Anthropic spent years warning that AI was too dangerous to release without safeguards. They positioned themselves as the responsible ones. They called for more regulation. They told us all how scared we should be.
They were right. It just turns out the danger was their own business partner.
