At 5:21 PM ET on June 12, the US government issued an export control directive ordering Anthropic to suspend all access to Claude Fable 5 and Mythos 5. Every user globally — not just foreign nationals — lost access immediately. Anthropic says it disagrees.
Context: This happened hours after Fable 5 was confirmed as the top-scoring model on FrontierMath Tier 4 — the hardest AI benchmark in existence — at 87% accuracy, surpassing OpenAI’s GPT-5.5 (72%).
According to Anthropic’s official statement, the US government cited national security authorities and a demonstrated “jailbreaking” method as justification for the recall.
The jailbreak in question: asking the model to read a specific codebase and fix any software flaws.
And here’s the part that makes this story extraordinary: the jailbreak research was conducted by researchers at Amazon — not a foreign adversary, not a rogue hacker. Amazon. A direct competitor in the AI race.
According to the Wall Street Journal, Amazon’s researchers used a series of prompts to get Fable 5 to provide information about a handful of security vulnerabilities. Anthropic shared the report with Katie Moussouris, chief executive of cybersecurity firm Luta Security and one of the most respected voices in infosec.
Her verdict was unambiguous:
Katie Moussouris, CEO, Luta Security
“Who at the White House evaluated this and thought it was a threat? It’s a complete overreaction because this is exactly the kind of prompting that defenders would do.”
She added that the information the model provided would be more useful to people defending computer networks than to those attacking them. Amazon declined to comment.
The Commerce Department issued the export control order. The DOJ is now considering whether to appeal a court order that could pause the ban.
Timeline — June 12, 2026
Morning
Epoch AI confirms Fable 5 scores 87% on FrontierMath Tier 4
Highest score ever recorded on research-level math
All Fable 5 and Mythos 5 access suspended globally
Evening
Anthropic publishes response: “We disagree.”
Commits to sharing more within 24 hours
Anthropic Pushes Back
In a rare public rebuke, Anthropic directly challenged the government’s rationale:
Anthropic Official Response
“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”
Anthropic made three pointed observations:
1
The vulnerability is “narrow and non-universal” — not a systemic safety risk
2
The same capability exists in OpenAI’s GPT-5.5 — which was not recalled
3
The demonstrated technique — reading code and finding bugs — is a core use case, not a jailbreak
Why Only Anthropic?
This is the question the AI industry is asking. GPT-5.5 can do the same thing. Google’s Gemini can do the same thing. Every frontier model can read code and identify vulnerabilities — it’s one of the most common enterprise use cases.
Fable 5 just demonstrated the highest reasoning scores ever recorded. When a model crosses a capability threshold, regulators notice.
Theory 2
Competitive Pressure
OpenAI has deeper government relationships. Anthropic — despite being founded by ex-OpenAI safety researchers — doesn’t have the same political insulation.
Theory 3
Precedent Setting
The government is establishing the mechanism for AI model recalls. Anthropic — the safety-focused lab — is the easiest first target because they’re least likely to resist legally.
The Structural Read
This is the Permission Layer in action — and it just became the most important layer in the Map of AI.
Until now, the AI race has been about capability: who builds the smartest model wins. This changes that equation. The government just demonstrated it can pull any model off the market, instantly, for any reason.
The New Equation
Capability x Permission = Deployable AI
You can build the best model in the world. If the government revokes permission, it doesn’t matter.
Three implications:
1. GOVERNMENT BECOMES KINGMAKER
If regulators can selectively recall models, they can pick winners. The lab with the best government relationships — not the best model — gets to stay deployed.
2. OPEN SOURCE GETS STRONGER
You can’t recall an open-source model. Meta’s Llama, Mistral’s models — once released, they can’t be pulled back. This event is the strongest argument yet for open-weight AI.
3. THE CHILLING EFFECT IS REAL
Every AI lab just watched the most safety-conscious company in the industry get its flagship model recalled. The message: building the best model makes you a target. That incentive structure is the opposite of what the US needs in an AI race with China.
Business Engineer Framework
The Permission Layer
The Map of AI has 9 layers. But there’s a layer above all of them that determines which models reach users: the Permission Layer. Government regulation, export controls, and safety mandates now control which AI capabilities are deployable — and which aren’t.
On the same day Claude Fable 5 was confirmed as the most capable reasoning model ever built, the US government pulled it from every user on Earth. Anthropic — the company that literally exists because its founders believed AI safety required a dedicated organization — just had its best model recalled over a capability that every competitor also has.
The AI race just got a new variable. It’s not just about who builds the best model. It’s about who’s allowed to deploy it.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
