Amazon Sabotaged Anthropic’s Fable 5 from inside — internal investigation reveals. Amazon orchestrated the operation that took down Fable 5. Andy Jassy escalated cyber concerns directly to the Treasury Secretary. Five other companies relayed the alarm to Washington the same evening. Anthropic refused a voluntary takedown, then complied under an export control order.
Key Takeaways
- Andy Jassy escalated an internal Amazon report on Fable 5 to the US Treasury
- Five other companies contacted the White House the same evening
- Anthropic refused a voluntary shutdown before being ordered to comply
The Amazon call that opened the file
Andy Jassy did not wait for a press release. The Amazon CEO personally escalated his team’s concerns to Treasury Secretary Scott Bessent and other senior government officials. The warning targeted Claude Fable 5, Anthropic’s flagship model released only days earlier.
Per Jassy’s account, Amazon researchers used Fable 5 to obtain information that could be used in cyberattacks. The Thursday evening before the federal order, Amazon delivered a formal report to Washington describing a technique to unlock the model. That document was the operational trigger for what followed.
The same Thursday evening, at least five other companies contacted senior government officials. The exact list has not been disclosed. Their move coincided with Amazon’s, and that simultaneity weighed heavily on the decision taken a few hours later.
Inside the White House, National Cyber Director Sean Cairncross pulled the relevant officials into a room. Hours of discussion followed, with a request for voluntary takedown sent to Anthropic. Dario Amodei and his team declined.
Four hours to take a model down
The clock tells the story. At 5:20 p.m. Eastern time, the administration issued an export control order targeting Fable 5 and Mythos 5. Anthropic had ninety minutes to comply. By 10:00 p.m. the same day, the company had shut down access to the model.
The speed is unprecedented. No US lab had ever been hit with this kind of federal directive against a commercially deployed AI product. Anthropic complied without public pushback. Compliance came before any public defense of the model.
Anthropic’s technical argument is only half convincing. The company maintains that the capabilities Amazon flagged are already available in other publicly accessible models. Per its statement, Fable 5 introduces no radically new cyber risk compared to the broader market.
Outside experts back this reading. Cybersecurity researcher Katie Moussouris called the technique Amazon flagged “Defense Oriented Prompting“, meaning a defensive usage pattern, not a jailbreak. Her take frames the government response as disproportionate to the actual risk.
Donald Trump’s former AI czar David Sacks gave a different version. According to him, a trusted partner revealed a jailbreak. The administration asked Dario Amodei to fix it or de-deploy the model. Amodei refused, which is what precipitated the federal order.
Also on Horizon:
- Meta AI Gateway: Bosworth cuts off internal tokenmaxxing
- OpenAI probe: state attorneys general turn up the heat
- Manus deal: Meta unwinds $2B under Beijing’s order
What this changes for Amazon, AWS and the AI ecosystem
In the short term, Amazon took a real commercial hit. AWS hosts Claude for its enterprise customers through Bedrock. The Fable 5 and Mythos 5 cutoff hit AWS users directly, which an Amazon spokesperson acknowledged. The group sacrificed part of its own product to take down Anthropic’s.
The strategic calculation reads between the lines. Amazon invests heavily in Anthropic but does not control its flagship model. By pulling Fable 5 and Mythos 5 off the market, the group brings the AI race down a notch, which gives more room for its own internal efforts to catch up. The competition between AI providers and their cloud hosts is now open and visible.
In the medium term, the precedent reshapes governance of frontier models. Any large customer can now trigger a federal process within hours against a provider whose model it considers too powerful. That lever worries rival labs, the same way the recent suspension recentered the AI sovereignty debate in India.
For Anthropic, the lesson is harsh. Refusing the voluntary takedown cost control of the calendar and the public narrative. The company is now explaining why its capabilities are equivalent to those of other authorized models, without being able to undo the decision. Its safety-first communication, leveraged as a competitive edge for months, just turned against it.
For the rest of the industry, the signal is direct. The largest customers of an AI lab can become its political opponents within forty-eight hours. Commercial contracts, investment ties, and cloud hosting deals no longer protect against an escalation to Washington. Labs will now arbitrate continuously between visible power and regulatory exposure.
Follow the story on Horizon.
