Trust, but it's monitored trust.
In the space between technological ambition and national security, Anthropic has navigated a two-week reckoning with the limits of trust. After federal authorities suspended access to Mythos 5 — a powerful cybersecurity AI — over fears of foreign infiltration and potential manipulation, the company worked daily with regulators to demonstrate that safeguards could be made real. Commerce Secretary Howard Lutnick ultimately agreed, restoring access to more than 100 critical infrastructure organizations in what may be less a resolution than a new kind of ongoing negotiation between those who build powerful tools and those who must answer for their consequences.
- A sudden mid-June shutdown cut off over 100 major corporations and government agencies from Mythos 5 after warnings that the model could be jailbroken and that a China-linked group had already gained access.
- The suspension exposed a fault line in AI deployment: even trusted domestic partners can lose access overnight when national security concerns override commercial agreements.
- Anthropic spent fourteen days in daily talks with federal officials, racing to design safeguards credible enough to satisfy a government that had already been burned once.
- Commerce Secretary Lutnick formally certified that 'appropriate safeguards are in place,' unlocking access for a defined set of trusted US partners — but leaving the broader public-facing model, Fable 5, still in limbo.
- Anthropic is now pushing to expand the approved user list and restore Fable 5, signaling that the current authorization is an opening move, not a final answer.
Anthropic has regained federal permission to share its most powerful cybersecurity AI, Mythos 5, with more than 100 American institutions — a hard-won outcome after two weeks of intensive negotiation with government regulators.
The crisis began on June 12, when Anthropic abruptly cut off all customer access to both Mythos 5 and its public-facing counterpart, Fable 5. The government had issued a directive to block access for all foreign nationals, including those working inside the United States. The order followed warnings from Amazon and others that the models could be jailbroken, and a more alarming report that a group with ties to China had already accessed Mythos 5.
Rather than accept the shutdown as permanent, Anthropic entered daily conversations with officials to design safeguards that could make access safe again. Commerce Secretary Howard Lutnick eventually concluded the company had done enough, writing in a formal letter that appropriate protections were now in place for certain trusted partners. Anthropic has also committed to ongoing collaboration with the government on protocols and future releases.
The restoration covers critical infrastructure operators and major institutions, but the story is not finished. Fable 5 remains suspended, with no timeline set for its return. Anthropic is also pressing to expand the approved user list beyond its current boundaries. What this episode reveals is how AI governance is actually unfolding — not through prewritten law, but through real-time negotiation, with security fears setting the tempo and trust being rebuilt one conversation at a time.
Anthropic has won back the right to share its most powerful cybersecurity tool with American institutions after two weeks of intensive negotiation with federal regulators. The company announced on X that the US government has approved the restoration of Mythos 5 access to what it described as "a set of US organizations that operate and defend critical infrastructure." According to reporting by Semafor, that set includes more than 100 institutions—major corporations and government agencies alike—whose access to the model had been completely cut off since mid-June.
The backstory here matters. On June 12, Anthropic made the abrupt decision to block all customer access to both Mythos 5 and Fable 5, a public-facing variant designed to bring similar capabilities to a broader audience. The company had received a directive from the government: suspend access for all foreign nationals, even those physically present in the United States and employed by Anthropic itself. The order came after Amazon and other companies warned federal authorities that the models could potentially be jailbroken—manipulated to behave in ways their creators never intended. There was also a more acute concern: US officials had been told that a group with ties to China had managed to gain access to Mythos 5.
Those were serious enough problems to warrant a complete shutdown. But they were also problems that Anthropic believed it could solve. Over the next fourteen days, the company engaged in daily conversations with government officials, working to design safeguards that would allow trusted American institutions to use the tool again without exposing it to the same risks. Commerce Secretary Howard Lutnick, who oversees these matters, eventually concluded that Anthropic had done the work. In a letter to the company, Lutnick wrote that he had "determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model." He also noted that Anthropic had committed to ongoing collaboration with the government on protocols, standards, and future releases.
The company is moving quickly to restore service to those 100-plus organizations. But this is not the end of the story. Anthropic is still in talks with the government about bringing back Fable 5, the model designed for broader public use. No timeline has been announced for that decision. The company is also pushing to expand the list of approved users even further, suggesting that the current authorization is a starting point rather than a final settlement.
What emerges from this sequence is a picture of how AI governance is actually happening in practice: not through legislation or clear rules written in advance, but through real-time negotiation between companies and regulators, with security concerns driving the pace and shape of access. Anthropic had to prove it could be trusted with a powerful tool. The government had to be convinced that the company's safeguards were genuine. Both sides appear to have found enough common ground to move forward—at least for now.
Citações Notáveis
I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.— Commerce Secretary Howard Lutnick
Anthropic has committed to work with the US government on protocols and standards and releases for its models.— Commerce Secretary Howard Lutnick
A Conversa do Hearth Outra perspectiva sobre a história
Why did the government order a complete shutdown in the first place? That seems extreme.
It was the combination of things. The jailbreaking risk was real—these models can be tricked into doing things they shouldn't. But the immediate trigger was the report that a China-linked group had actually gotten into Mythos 5. That's not theoretical. That's an active threat.
So Anthropic had to prove it could keep bad actors out. How do you even do that?
You build walls. You verify who's accessing the system. You limit it to organizations you can actually vet and monitor. That's why it's only going to critical infrastructure operators and government agencies right now—people with their own security infrastructure, people the government already knows.
But Anthropic wants to release Fable 5 to the public eventually. How does that work if the underlying technology is still vulnerable?
That's the real tension. You can't jailbreak what you don't have access to. So the public version would have to be different—weaker, more constrained, harder to manipulate. Or the safeguards have to be so good that jailbreaking becomes impractical. Anthropic is betting on the latter.
And the government is just... trusting them on that?
Not exactly. They're trusting them with limited access while they watch. This approval is conditional. Anthropic has to keep talking to Commerce, keep reporting, keep proving the safeguards work. It's trust, but it's monitored trust.
What happens if there's another breach?
Then you're back to square one. The shutdown happens again. The whole arrangement depends on Anthropic actually delivering on what it promised.