The government's hand remains on the switch.
Within weeks of their debut, two of Anthropic's most capable AI models were pulled from global circulation by the US government — a rare and striking act that placed the question of who controls powerful technology squarely at the center of public life. The ban has now been lifted, but not without cost: Anthropic has accepted a form of ongoing government oversight that redraws the boundary between private innovation and public accountability. What emerges from this episode is less a resolution than a new arrangement — one in which advanced AI capability is understood, at last, as something that carries sovereign weight.
- Two of Anthropic's most powerful AI models were abruptly pulled from global markets within days of launch after a jailbreaking vulnerability exposed their potential for misuse.
- The government's swift action signaled that it views frontier AI not merely as a commercial product but as a potential instrument of national harm — a designation with serious consequences for the industry.
- Anthropic pushed back, arguing the vulnerability was narrow and the recall disproportionate, but ultimately negotiated rather than fought — accepting proactive security monitoring, advance government consultation on future releases, and mandatory reporting of malicious activity.
- The ban has been lifted, but the Commerce Department explicitly retained the right to reinstate restrictions, leaving Anthropic operating under a conditional truce rather than a clean restoration.
- The episode sets a precedent: advanced AI companies may now face a regulatory posture in which the government holds a standing switch — ready to act, and willing to use it.
In mid-June, the US government ordered Anthropic to withdraw Claude Fable 5 and Mythos 5 from global circulation. The two models had launched on June 9th; within days, they were gone. The trigger was a discovered jailbreak in Fable 5 — a method for bypassing its safety guardrails and unlocking capabilities the company had deliberately restricted. The Commerce Department moved quickly, and the suspension held despite Anthropic's objections that a single narrow vulnerability did not warrant a full recall.
Now, less than three weeks later, the ban has been lifted. Commerce Secretary Howard Lutnick confirmed the decision in a letter to the company, though the government did not publicly detail what specific fixes had satisfied its concerns. What Anthropic did agree to was substantial: proactive hunting for security vulnerabilities before models reach users, advance government consultation on future releases, and mandatory reporting of any detected malicious activity. These commitments amount to a new and ongoing form of oversight — closer to a regulatory partnership than a typical company-regulator relationship.
The two models occupy distinct markets. Fable 5 is a consumer-facing reasoning engine capable of complex, multi-step tasks. Mythos 5 targets businesses and cybersecurity professionals, and its sensitivity stems from a specific capability: it can scan code for vulnerabilities and exploit them. That combination made it a particular concern from a national security standpoint.
The restoration of access is a win for Anthropic, but a conditional one. The Commerce Department explicitly reserved the right to reimpose restrictions if new risks emerge — meaning this is a truce, not a resolution. The government has demonstrated both the willingness and the speed to act, and it has made clear that its hand remains on the switch.
In mid-June, the US government ordered Anthropic to pull two of its most powerful artificial intelligence models from the market. Claude Fable 5 and Mythos 5 vanished from circulation almost overnight, suspended worldwide over fears that hackers could weaponize them. The models had just launched on June 9th. Within days, they were gone.
The government's concern was specific: someone had found a way to jailbreak Fable 5, to slip past its safety guardrails and unlock capabilities the company had deliberately restricted. Anthropic pushed back at the time, arguing that a single narrow vulnerability in a model used by hundreds of millions of people did not justify a full recall. But the Commerce Department disagreed. The suspension held.
Now, less than three weeks later, the ban has been lifted. Anthropic announced this week that it will begin restoring access to both models after receiving word from the Department of Commerce that the restrictions have been removed. Commerce Secretary Howard Lutnick wrote to the company confirming the decision, though the government did not publicly detail what specific fixes had satisfied its concerns.
What Anthropic did agree to, according to Lutnick's letter, was a new framework for managing risk. The company committed to actively hunting for security vulnerabilities in its models before they reach users. It also agreed to work directly with the government on future releases, giving officials advance notice and input before new versions go live. And it promised to report any malicious activity it detects to federal authorities. These were not minor concessions. They amount to a form of ongoing oversight that extends beyond the typical relationship between a private company and a regulator.
The two models at the center of this episode represent different markets. Fable 5 is built for consumers—a reasoning engine capable of handling complex, multi-step tasks with minimal human intervention. Mythos 5 targets businesses and cybersecurity professionals, and it carries particular weight because of what it can do: it can scan computer code for vulnerabilities and, crucially, it can exploit them. That capability made it especially sensitive from a national security standpoint. In the wrong hands, a tool that finds and weaponizes security flaws could cause real damage.
Anthropicargued at the time that the government had never spelled out its specific technical objections, that the company was being asked to solve a problem that was never clearly defined. But the company also understood the stakes. An export ban on advanced AI models is not a minor regulatory action—it signals that the government views the technology as a potential national security threat, something that cannot be allowed to circulate freely. The company negotiated its way back into the market by accepting a tighter leash.
The Commerce Department, notably, reserved the right to reimpose restrictions if new risks emerge. That clause matters. It means this is not a permanent resolution but a conditional one, a truce that holds only as long as Anthropic maintains the security posture it has now promised. The government is watching, and it has shown it is willing to act quickly if it believes the risk has returned. For Anthropic, the restoration of access is a win—but it comes with the understanding that the government's hand remains on the switch.
Citações Notáveis
Anthropic has agreed to proactively detect and address security risks associated with the models.— Commerce Secretary Howard Lutnick
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, in statement during suspension
A Conversa do Hearth Outra perspectiva sobre a história
Why did the government move so fast to ban these models in the first place? Was there an actual attack, or just the possibility of one?
Just the possibility. Someone found a jailbreak—a way to bypass the safety features—but there's no evidence it was used maliciously. The government acted on the vulnerability itself, not on harm that had already occurred.
And Anthropic disagreed with that decision?
They did. They said a narrow flaw in a model used by hundreds of millions of people didn't justify pulling it entirely. But they also understood they didn't have much leverage. A national security suspension is serious.
So what changed in three weeks? Did they actually fix the jailbreak?
The government's letter doesn't say. It just says Anthropic addressed the risks. What we know they agreed to is different—proactive monitoring, government collaboration on new releases, reporting malicious activity. It's less about the technical fix and more about oversight.
That sounds like the government got what it wanted without necessarily solving the original problem.
Possibly. But it also means the government now has visibility into what Anthropic is building before it ships. That's a different kind of security—not preventing the flaw, but catching it early and having a partner who reports what they find.
And if something goes wrong again?
The Commerce Department kept the right to reimpose the ban. This isn't permanent. It's conditional on Anthropic holding up its end of the agreement.