The lights went out for Mythos 5 and Fable this weekend.
Anthropic pulled the plug. Access vanished.
Why?
The federal government told them Friday afternoon they had found a way to jailbreak Fable 5.
It was a bypass. A hole in the armor. One that let users dodge the limits Anthropic had baked in to stop misuse.
Remember when Anthropic launched Mythos?
They didn’t open the floodgates. Not really. They gave it to a select few. Government agencies. Tech pros. Cybersecurity nerds. Because the thing could spot vulnerabilities like a hawk spotting a mouse.
Now?
Export controls. The scary kind.
The US government said the jailbreak was a problem. A big one. So Anthropic had to lock out anyone who isn’t a US person.
Foreign nationals. Out.
Whether they are in America or not. Doesn’t matter.
The catch? Anthropic can’t easily separate users. They can’t just filter out the tourists or the expats sitting in San Francisco. So they did the nuclear option.
They shut the whole thing down.
Late Friday, Anthropic pushed back. They disagreed with the severity of the response. They argued that if this standard stuck—if the feds banned a model every time someone found a trick to break its safety rules—it would essentially halt frontier AI research. Period.
“We believe the government should have the authority to block unsafe deployments… through a process that is transparent, fair… and grounded in technical facts. This action does not adhere to those principles.”
A statement is nice. It doesn’t open the API.
Anthropic sent a team to Washington. Talks were held. Calls were made. On Saturday, company leaders spent hours on the phone with Commerce Secretary Howard Lutnick and National Cyber Dir Sean Cairncross. The WSJ reported the huddle.
It’s not their first run-in with D.C.
Earlier this year, the Pentagon declared Anthropic a supply chain risk. Why? Because Anthropic wanted guardrails. No fully autonomous weapons. No mass domestic surveillance.
The Pentagon wanted free reign. For all lawful purposes.
David Sacks? He didn’t help. The former crypto czar, now sitting on a presidential science council, took to X (Twitter). He dismissed any link between the Pentagon feud and this shutdown. But he did say the ball was in Anthropic’s court.
Sacks argued Anthropic put profits before safety. They kept a consumer model running while a security flaw existed.
Easy for him to say.
Ayham Boucher, an AI strategist at Cornell, thinks this is much harder than a patch note.
All models can be jailbroken.
That’s just the nature of the beast. A jailbreak is a prompt injection. A roleplay request. Tell the AI you’re writing a screenplay. Suddenly, it tells you how to rob a bank because it’s not “committing” the crime. It’s “fiction.”
As models get smarter, they get better at these tricks.
Anthropic tried. For Fable 5, they stored chat data for 30 days. Unusual. Some businesses hated it. Data privacy is a sensitive spot. But the storage helped Anthropic catch bad patterns. Find the holes. Plug them.
The government wants Anthropic to fix this hole.
Boucher says the government misunderstands the tech. There will always be another hole. Always.
“Mythos” sounds exclusive. Rare.
But it isn’t. Not really.
Other labs can do it. OpenAI’s GPT-5.5 is already close in cybersecurity skills. Maybe not identical, but the gap is closing fast.
Why? Because math and code scale beautifully. Writing is messy. Hard to verify objectively.
Math is binary. Code is logic. These are easy for models to learn. And they are learning exponentially.
This isn’t just Anthropic vs. the world.
It’s the entire field. Chinese devs at DeepSeek or Alibaba are coming. No roadblocks for them in coding skills. Just data and compute.
The Trump administration talks about keeping America ahead of China. They tout deregulation as the key to global dominance.
So why pull the plug now?
Maybe Mythos jolted them.
It shows a messy pivot in DC. Last year, the plan was: limit regs to boost dominance. Now? President Trump signed a voluntary exec order. Let the government review frontier models before they launch.
Voluntary?
“The voluntariness of that process seems lasting only a short time,” said Samir Jain of the Center for Democracy and Technology, “The voluntariness seems to have lasted a short time.”
Wait. The quote repeats in the source but implies brevity.
Let’s fix that logic. The voluntary nature lasted seconds.
Now we have opaque actions. Behind closed doors.
No clear process. No clear reasoning. Just a ban.
Jain points to the First Amendment.
AI generates speech. Developers make editorial choices. That’s speech. Regulating it? You need the rule of law.
You need a clear interest. Compelling.
But this? This feels arbitrary.
Does it secure the country?
Or does it slow down the engine?
Boucher says it’s not just the users banned. It’s the talent. Foreign nationals working at Anthropic are locked out of their own tools. They can’t help build the next generation of the model they help design.
“I can’t see how this isn’t slowing down model development,” he said.
Washington wants security.
Developers want freedom to ship.
Somewhere in between, a model was turned off. And no one knows if it will ever come back on.
Or what standard applies next.
