The US government just drew a line between AI that defends and AI that could attack, and only one of Anthropic's flagship models made the cut.
The Summary
- Anthropic regained US approval for Mythos 5, its cybersecurity-focused AI model, after a two-week lockdown triggered by export control concerns under the Trump administration
- Fable 5 remains restricted, with no timeline for restoration, revealing a government willingness to greenlight defensive AI while keeping general-purpose frontier models locked down
- Access is limited to US organizations operating critical infrastructure, not a blanket approval
- The restrictions barred foreign nationals, including Anthropic's own employees, from accessing both models starting June 12
The Signal
On June 12, Anthropic disabled access to both Mythos 5 and Fable 5 in response to a US export control order. The order didn't just block foreign users. It blocked foreign national employees inside Anthropic's San Francisco headquarters from touching the models. For two weeks, CEO Dario Amodei negotiated with the White House while two of his company's most powerful models sat dark.
Friday brought partial relief. The government cleared Mythos 5 for "a set of US organizations that operate and defend critical infrastructure," Anthropic said. Not everyone. Not even all US companies. Just the ones running power grids, water systems, and financial networks. Fable 5 stayed locked.
"The US is prioritizing national security while still enabling domestic AI adoption in critical sectors."
The distinction matters. Mythos 5 is Anthropic's cybersecurity specialist. It finds vulnerabilities, hardens systems, defends networks. Fable 5 is the general-purpose reasoning model, the one built for open-ended problem solving. The government looked at both and said: the defensive tool can come back online under strict conditions. The reasoning engine stays dark until we figure out what you might build with it.
This isn't about capability limits or safety benchmarks. Both models cleared Anthropic's internal red-teaming. The concern was "potential threats to national security," which in policy language means dual-use risk. An AI that writes exploit code is also an AI that could teach adversaries how to write exploit code. An AI that reasons through complex strategic problems could reason through problems you'd rather it didn't.
Key dynamics at play:
- The US is carving AI policy by use case, not by capability threshold alone
- Access geography now includes citizenship of employees, not just end users
- "Critical infrastructure" becomes the wedge category for frontier AI deployment
Former White House CIO Theresa Payton framed the approval as evidence the US wants domestic AI adoption in key sectors without ceding ground on security. But the Fable 5 lockdown tells a different story. The government isn't just gatekeeping foreign access. It's deciding which problems advanced AI should be allowed to solve, and who gets to deploy it.
Anthropic says it's "continuing to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again." No timeline. No clarity on what would need to change for Fable 5 to come back. Just negotiation in progress while one of the world's most capable reasoning models sits behind a policy wall.
The Implication
If you're building on frontier models, pay attention to capability class, not just capability level. The government just signaled it will approve narrow, defensive AI tools faster than general-purpose reasoning systems, even from the same company with the same safety protocols. Cybersecurity agents, fraud detection, infrastructure monitoring? Likely green-lighted. Strategic planning, open-ended analysis, general automation? Expect scrutiny.
For infrastructure operators, this is your window. Mythos 5 access is live for critical sectors now. If you're defending networks, managing grid systems, or running financial rails, you have a narrow advantage while competitors wait for broader model access. Use it.