One of America's champions is being kneecapped by the US government
In the span of a single Friday afternoon, the Trump administration handed Anthropic a ninety-minute ultimatum to disable its most advanced AI models, forcing a company that had spent years navigating the tension between capability and caution into an overnight crisis of a different kind. The directive, rooted in reported jailbreak vulnerabilities and the specter of foreign exploitation, arrived not as policy deliberated in the open but as a command delivered under pressure — a reminder that in the age of strategic technology, the state retains the power to unmake what the market has made. What unfolded over the following weekend was less a resolution than a reckoning: with the limits of regulatory frameworks, with the meaning of competitive advantage, and with the question of whether safety and sovereignty can be pursued without undermining both.
- Anthropic received a 90-minute ultimatum on a Friday afternoon to shut down Mythos 5 and Fable 5 globally or face forced export controls — leaving the company no viable path but compliance.
- CEO Dario Amodei was on calls with the White House within fifteen minutes, and by the weekend, senior executives were flying to Washington for emergency face-to-face negotiations with Treasury, Commerce, and the National Cyber Director.
- The alleged vulnerability — a jailbreak method Anthropic says is narrow, non-universal, and present in competitors' systems including OpenAI's GPT-5.5 — has left the industry questioning whether the directive targets a real threat or sets a dangerous precedent.
- Tech and cybersecurity leaders organized a public letter over the weekend warning that the ban hands adversaries a strategic gift, with one executive stating bluntly, 'They are laughing at us in Beijing right now.'
- Across the industry, companies are already signing backup contracts with non-US providers and treating political risk as a formal line item — signaling that the damage to business confidence may outlast any regulatory resolution.
On a Friday afternoon in June, while much of America was absorbed in World Cup celebrations and a Knicks championship, Anthropic received a call that would consume the next seventy-two hours. The Trump administration gave the San Francisco AI company ninety minutes to cut off foreign access to Mythos 5 and Fable 5 — two of its most powerful models — or face Commerce Department export controls imposed by force. By early evening, the directive was in writing. Anthropic shut down products it had spent the previous week promoting to the world.
Mythos 5 and Fable 5 sit at the frontier of AI-assisted cybersecurity. Mythos Preview, the foundation model, had been deemed too sensitive for public release. Mythos 5 went to select government and enterprise clients. Fable 5, equipped with additional safeguards, was cleared for general use — safeguards so stringent they had become a running joke in the cybersecurity community on launch day. A report surfaced suggesting those guardrails could be bypassed, and the administration moved swiftly.
Dario Amodei was on calls with the White House within fifteen minutes of first contact. Over the weekend, he spoke directly with Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and National Cyber Director Sean Cairncross. Anthropic flew its head of safeguards, its frontier red team leader, and a senior cybersecurity researcher to Washington for in-person meetings. The company acknowledged the jailbreak concern but called it narrow and non-universal — and noted pointedly that the same capability existed in OpenAI's GPT-5.5 and other competitors' systems.
The origins of the alarm remain murky. Some reports pointed to a China-linked group accessing the technology; others credited Amazon researchers who had red-teamed Fable 5 and surfaced vulnerabilities. What seemed clear was that Anthropic had pre-briefed the administration on Fable 5, and the Commerce Department had conducted its own testing before deployment without raising objections.
The industry response was immediate. A public letter organized by Alex Stamos of Corridor argued that AI systems are essential for patching vulnerabilities before adversaries exploit them — a race, he said, where the United States leads by only months. Ben Van Roo of Legion Intelligence called the directive 'the most impossible thing to enforce,' noting the legal and practical maze of blocking foreign nationals from a model available online. The stakes extend well beyond Anthropic: if the administration bans these models, it can justify banning comparable offerings from OpenAI, Google, and Microsoft.
By Monday, no resolution had emerged. Companies across the industry were already signing backup contracts with non-US providers and deploying open-weight models on alternative hardware. Political risk had become a formal line item in strategic planning. What happens next will reveal whether the administration views AI safety as a genuine national security priority — or whether competitive pressure from China will force a recalibration.
On a Friday afternoon in June, while much of America was celebrating a World Cup victory and the New York Knicks championship, Anthropic received a phone call that would consume the next seventy-two hours. At 1 p.m. Eastern time, the Trump administration gave the San Francisco AI company ninety minutes to disable access to two of its most powerful models—Mythos 5 and Fable 5—to any foreign national, including its own employees from abroad. If Anthropic refused, the Commerce Department would impose export controls by force. By 5:21 p.m., the directive arrived in writing. The company's only option was to shut down products it had spent the previous week promoting to the world.
Mythos 5 and Fable 5 represent the cutting edge of Anthropic's work in AI-assisted cybersecurity. Mythos Preview, their foundation model, had been deemed too risky for public release—a judgment Anthropic had broadcast loudly, whether as genuine caution or strategic marketing or both. Mythos 5 went to select government agencies and companies. Fable 5, equipped with additional safeguards, was cleared for general use. But a report surfaced suggesting those guardrails could be bypassed. The administration, acting on that intelligence, moved swiftly.
AnthropicCEO Dario Amodei was on calls with the White House within fifteen minutes of that initial contact. Over the weekend, he spoke directly with Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and National Cyber Director Sean Cairncross, sometimes multiple times. The company flew its head of safeguards, Dave Orr; its frontier red team leader, Logan Graham; and cybersecurity researcher Nicholas Carlini to Washington for in-person meetings. Anthropic's public statement acknowledged the government's concern about a jailbreak method but characterized it as narrow and non-universal—something the company said it had voluntarily shared with authorities. More pointedly, Anthropic noted that the same capability was widely available in competitors' systems, including OpenAI's GPT-5.5.
The origins of the alarm remain partially obscured. Reports suggested a China-linked group had accessed the technology, though Anthropic said it had immediately revoked access to a major telecommunications company once the government flagged concerns. Other accounts pointed to Amazon CEO Andy Jassy, whose researchers had red-teamed Fable 5 and discovered vulnerabilities. David Sacks, the administration's former AI and crypto czar, credited a "highly credible trusted partner" for surfacing the jailbreak. What seemed clear was that Anthropic had pre-briefed the administration on Fable 5, and the Commerce Department had conducted testing before deployment with no objections raised at the time.
The industry response was swift and sharp. On Sunday, a public letter from tech and cybersecurity executives called for the restrictions to be lifted. Alex Stamos, chief product officer at Corridor, organized the effort, arguing that AI systems are essential for patching vulnerabilities before adversaries exploit them—a race, he said, where the United States leads by only months. "They are laughing at us in Beijing right now," Stamos told The Verge. "One of America's champions is being kneecapped by the US government while we're in a race with the Chinese." The letter noted that Fable 5's safeguards were so stringent they had become a source of humor in the cybersecurity community on launch day, and that Mythos-class models, while skilled at finding exploits, were not uniquely good at the task.
Ben Van Roo, CEO of Legion Intelligence, a system serving the national security community, called the directive "the most impossible thing to enforce." Preventing any foreign national from accessing a model available online creates a practical and legal maze. The stakes extend beyond Anthropic. OpenAI, Google, and Microsoft have released comparable products with similar claims about their power and risk. If the Trump administration bans Anthropic's models, it can justify banning theirs. The weekend's conversations between Anthropic and the administration were described as constructive, with some officials acknowledging that export controls on model providers create problems—especially when competitors offer the same capabilities and the government is simultaneously trying to encourage American AI exports.
By Monday, no resolution had emerged. Anthropic's negotiations continued in what Van Roo called "uncharted territory," a regulatory space no one had navigated before. The company also faces a separate, ongoing battle with the Pentagon over how the military can use its technology. The weekend's events have already reshaped business calculations across the industry. Companies are signing backup contracts with non-US providers and deploying open-weight models on alternative hardware. Political risk is now a line item in strategic planning. What happens next will signal whether the Trump administration views AI safety as a genuine national security concern or whether it will recalibrate its approach in light of competitive pressure from China.
Citações Notáveis
We're in a race, and I think policymakers don't understand that. If the labs are ahead, it's only by a matter of months.— Alex Stamos, chief product officer at Corridor
The directive of 'no foreign national should use this model' is the most impossible thing to enforce.— Ben Van Roo, CEO of Legion Intelligence
A Conversa do Hearth Outra perspectiva sobre a história
Why did the administration move so fast? Ninety minutes seems almost panicked.
They were responding to a report of a vulnerability in a model they'd already cleared. Once that alarm was raised, they had to act—or appear negligent if something went wrong. The speed was partly about control, partly about showing they take the threat seriously.
But Anthropic says the same vulnerability exists in OpenAI's models. So why only target Anthropic?
That's the question everyone's asking. Anthropic had been more vocal about the risks, had made bigger claims about Mythos being dangerous. When those claims came back to haunt them, the administration had leverage. It's also possible they're using Anthropic as a test case—seeing what they can enforce before moving on competitors.
The cybersecurity people seem furious. Why?
Because they see this as self-sabotage. These models are tools for finding and fixing vulnerabilities before bad actors exploit them. Banning them doesn't make America safer—it just means American companies can't use the best tools available. Meanwhile, China keeps developing its own systems.
Is there a real security risk here, or is this political theater?
Probably both. The jailbreak is real—someone found a way to bypass safeguards. But Anthropic's point stands: the capability isn't unique to their model. The question is whether the administration is genuinely concerned about that specific vulnerability or whether they're using it as cover for broader concerns about AI export and control.
What does Anthropic lose if this sticks?
Everything they've built toward. Mythos was supposed to help them recover from months of fights with the Pentagon. Now their flagship product is disabled. They lose revenue, credibility, momentum. And if competitors face the same restrictions, the entire American AI industry loses leverage in a race they're already worried about losing.
Will this happen again?
Almost certainly. This is the first time a model's been disabled by government order. It won't be the last. Every company is now calculating the political risk of releasing powerful AI. That changes how they build, what they release, where they operate.