The Call Came From Inside the Cap Table

When I wrote that the US government export-controlled Fable 5 off the planet three days after launch, the trigger was a faceless “another company” that had jailbroken the model. The face has a name now, and it is the last one Anthropic wanted: Amazon, its single biggest investor.

Per Axios, Amazon researchers jailbroke Fable 5, wrote it up, and called administration officials on Thursday night. Andy Jassy escalated to Treasury Secretary Scott Bessent and the White House. Trump signed off. The wolf Anthropic spent fourteen months selling Washington didn’t come from an adversary. It came from inside the cap table.

The Mechanics of a Friday-Night Takedown

The timeline is brutal and worth seeing whole:

• Thursday night: Amazon’s report lands with senior officials. At least five other companies call too, through Thursday evening and Friday morning. A coordinated pile-on, not a lone whistleblower.
• Friday 1pm ET: Anthropic gets a call. Ninety minutes to take Fable and Mythos down for a “national security threat.” No details provided.
• Friday 5:30pm: the White House letter lands, export controls attached.
• Friday 10pm: users worldwide lose access.

The detail that should sting: Anthropic had notified the government multiple times about the June 9 launch, and nobody objected. The objection arrived only after a competitor-slash-investor put it on the phone.

It raises questions about why Amazon would strike such a disruptive blow against a company in which it is a major investor.

— Axios, June 13

And the timing is the wound. Anthropic filed its S-1 days earlier at a valuation past OpenAI, with the most capable model it had ever shipped as the headline asset. The takedown didn’t just dark a product for a few weeks. It put “the US government can switch our flagship off on a Friday” into the risk section of a prospectus that hundreds of millions of dollars are being priced against. A rival could not have aimed the blow more precisely if it tried.

The “Jailbreak” Was the Sales Brochure

Here is the part that turns this from drama into farce. What did Amazon’s researchers actually do? They prompted the model to find software vulnerabilities and write proof-of-concept scripts, tested against known CVEs and code they had deliberately broken themselves. No novel vulnerabilities in real-world software. Nothing an attacker couldn’t already get.

That is not a jailbreak. That is the Project Glasswing demo. It is the exact capability Anthropic put on a pedestal in April: a model that reads a codebase and finds the flaws, marketed as the future of cyber defense. Amazon ran the hero feature, relabeled the output “national security threat,” and called it in.

Defenders need to use the latest AI models in exactly this way to keep up with attackers.

— Katie Moussouris, CEO of Luta Security

She called the research narrowly focused, surfacing no new real-world risk, and the government’s response overreach. She is right, and Anthropic agrees: it told officials the same jailbreak works on other public models including GPT-5.5, which carry no such controls.

What This Isn’t

The betrayal framing is satisfying and I want to bound it honestly:

• It wasn’t only Amazon. Five-plus companies called. That makes this structural, not a single grudge: an industry discovering it can route a rival’s launch through a national-security panic button. That should worry everyone, Amazon included.
• The cloud-provider defense is real. AWS does field government security questions constantly. Reporting a capability is not automatically sabotage, even when you own equity on both sides.
• The ban will likely lift. An official said weeks. The cap table survives. What doesn’t survive is the discovery that it can be done at all.

The Loaded Gun Was Always Pointed Both Ways

I keep returning to one thing: none of this works if Anthropic hadn’t built the weapon narrative first. For over a year the pitch was that frontier models are so dangerous they are a national-security matter. That story was the moat, the safety-lab brand, the thing that justified the velvet rope. It was also a loaded gun left on the table.

This week a co-owner of the company picked it up. The capability didn’t change between Anthropic’s heroic Glasswing framing and Amazon’s national-security-threat framing. Only the narrator changed, and the narrator decides whether a model that finds bugs is a fire department or a bomb.

You cannot spend fourteen months teaching Washington that your product is a weapon and then act betrayed when the people closest to you hand it to the regulators. The call didn’t come from China. It came from the cap table, and the cap table knew exactly where the trigger was, because Anthropic drew the map.