Amazon owns a slice of Anthropic. The WSJ says its CEO helped get Fable 5 switched off

Who actually flipped the switch on Fable 5

On Friday, June 13, the U.S. government directed Anthropic to disable Fable 5 and Mythos 5 for foreign nationals — a restriction so broad it could not be cleanly scoped, so Anthropic pulled access outright. The Wall Street Journal's account (paywalled) names the trigger: conversations between Amazon CEO Andy Jassy and U.S. officials. According to people the paper spoke to, Amazon researchers had chained prompts to coax Fable 5 into producing information usable in cyberattacks, and that demonstration reached Washington.

The piece rests on single-thread "people familiar with the matter" sourcing, which several readers in the discussion flagged as thin. The fact the article doesn't dwell on is the one the thread couldn't stop turning over: Amazon owns more than 5% of Anthropic and, through AWS, supplies the compute its models run on. The company that demonstrated the jailbreak is an investor in, and landlord to, the company it reported.

A shareholder's disclosure, routed through the government

If you find a way to break a model, the normal path is responsible disclosure to the lab. Going to federal officials instead is a different act, and from a major shareholder a stranger one. Not everyone read malice into it:

"AWS is a partner on project Glasswing… So I don't think there is anything sinister here, I would use Hanlon's razor." — the co-founder and CTO of supply-chain security vendor Arnica

The most concrete business theory had nothing to do with security. It pinned the rupture on a contract fight over where customer data lives:

"Anthropic really ruined their Bedrock story by imposing data retention requirements that cross a red line in regulatory compliance… Jassy would rather have nobody use Fable than doing it on the basis of, effectively, a direct data trust relationship with Anthropic." — a commenter with roughly 20k karma walking through the Bedrock mechanics

A third read inverted the headline's arrow of causation, casting the government as initiator and Amazon as cover:

"I don't think this is Amazon targeting Anthropic, but the government shaking down Anthropic using Amazon. The government is a key customer of Amazon, so Amazon will provide cover as needed." — a reader reframing who pressured whom

The premise one tester says doesn't hold

The whole case assumes Fable 5 is uniquely dangerous to jailbreak. One commenter claiming hands-on red-team work argued the opposite: that the model resists exploitation even after you get in.

"fable is trained in a way that even if you were to jailbreak it, it would be completely uninterested in exploitation… Opus 4.8 is far more powerful in that regard." — a tester who reported spending $6,000 over three days pushing success rates from 3% to 85% on security tasks

That is an unverifiable self-report. But it points at the load-bearing weakness of the official story: the "produce attack info" demo can be run, with effort, against most frontier models. Singling out the one whose largest cloud partner happened to bring the demo to officials is the part that doesn't reduce to safety.

Why the circular money is the real story

Strip out the personalities and what remains is structural. Anthropic's compute, a slice of its equity, and now its regulatory exposure all route through one company that also competes with it. When a model can be switched off for "foreign nationals" on the strength of a vendor's demonstration and a phone call — no lawsuit, no published vulnerability, no court — the lesson isn't about one weekend. It is that the tangled cross-ownership of the AI industry hands any sufficiently large partner a lever, and "security" is the easiest label to bolt onto it.