The US government just drew a line between AI models that defend America and AI models that might teach it things it doesn't want taught.
The Summary
- Anthropic got partial approval to restore access to Mythos 5, its cybersecurity-focused AI model, after two weeks of lockdown under export controls
- The carveout only covers US organizations operating critical infrastructure, not general access
- Fable 5, Anthropic's other frontier model, remains completely restricted
- The original June 12 order blocked foreign nationals, including Anthropic's own foreign employees, from accessing both models
The Signal
On June 12, Anthropic disabled access to its two most powerful models after receiving an export-control order from the Trump administration. The order didn't just block foreign users. It blocked foreign national employees inside Anthropic itself, effectively quarantining the company's best work from chunks of its own engineering team.
Two weeks later, the company got a partial reprieve. Mythos 5 can now be deployed to US organizations that operate and defend critical infrastructure. Power grids, water systems, financial networks. The things you don't want going down. The White House is essentially saying: we trust this model to protect American infrastructure, but we don't trust it in general circulation.
"The government is treating different AI capabilities as different threat classes, and Mythos 5 passed the test that Fable 5 didn't."
Fable 5 remains locked. Anthropic hasn't disclosed what makes Fable different from Mythos, but the government's divergent treatment tells the story. One model gets cleared for defensive work on critical systems. The other stays in the vault. The implication: Fable 5 has capabilities the administration considers too risky for deployment, even in restricted form.
Key differences between the models:
- Mythos 5: cybersecurity-focused, now approved for critical infrastructure defense
- Fable 5: capabilities unknown, remains fully restricted
- Both were frontier models significant enough to trigger export controls
This marks the first time we've seen the US government carve up AI model access by use case rather than user nationality. Previous export controls focused on who could touch the technology. This focuses on what the technology can do and where it gets pointed. Bloomberg notes the approval came after Anthropic resolved "concerns about the technology's potential threats to national security", but those concerns clearly weren't uniform across models.
The timing matters. Anthropic CEO Dario Amodei has been in active negotiations with the White House throughout this period. The company is trying to expand access to Mythos 5 beyond critical infrastructure operators and get Fable 5 back online entirely. But the government now has proof of concept that it can bifurcate AI deployments based on capability profiles.
The Implication
If you're building frontier AI models, the regulatory calculus just changed. The US government will evaluate your models individually, not as a product suite. A company can have one model cleared for sensitive infrastructure work while another model from the same lab stays locked down. Specialization might become a regulatory advantage. A model purpose-built for cybersecurity defense carries different risk than a general-purpose reasoning model.
Watch how other labs respond. If Anthropic gets Fable 5 released, the criteria the government used to approve it will become the template. If Fable stays restricted, expect more labs to build use-case-specific models rather than chase general capability. The path of least regulatory resistance just became the path of most specificity.