US Orders Anthropic to Geofence Fable 5 and Mythos 5; Models Pulled Globally
Washington cites a jailbreak risk. Anthropic disagrees but complies, suspending both models worldwide rather than build a foreign-national access wall.
The US government has directed Anthropic to block all foreign nationals from accessing its Fable 5 and Mythos 5 models. Rather than stand up the access controls that order would require, Anthropic has suspended both models globally.
The company is complying. It is also publicly disputing the rationale.
In a statement, Anthropic characterized the jailbreak cited by the government as narrow, and argued the underlying capability is already widely available across other frontier and open-weight systems. That framing matters. It reframes the order as a policy signal about who gets access to frontier models, not a technical containment measure tied to a unique capability.
The directive appears to fall under export-control-style authorities that treat model access by non-US persons as a controlled transaction. Neither the agency issuing the order nor the specific legal instrument has been formally published at time of writing, and I'd treat single-source characterizations of the underlying jailbreak with appropriate skepticism until a primary advisory lands.
What we can say with medium confidence:
- The scope is access-based, not weights-based. There is no indication of a weights seizure or training-data action.
- The trigger was a specific capability concern, not a general policy review. Anthropic's own language ("narrow") implies a discrete elicitation, not a class of misuse.
The operational picture for defenders and red teams is messy. Customers building on Fable 5 or Mythos 5 — including non-US subsidiaries of US companies — are offline until Anthropic implements nationality-gated access or the order is modified. Enterprises with mixed-nationality engineering teams will feel this fastest. Expect short-term migration to Claude's prior generation, GPT-class models, or Gemini, depending on workload.
The attribution question worth watching is which jailbreak. Public red-team writeups over the past quarter have flagged several elicitation techniques affecting frontier reasoning models, with overlapping TTPs across labs. Without the government naming the capability, outside analysts are guessing.
A few threads to pull on:
- Whether other US labs receive parallel orders. If Fable 5 and Mythos 5 were singled out, the capability is model-specific. If OpenAI, Google DeepMind, and xAI get similar letters, it's a class action against frontier access.
- Whether Anthropic's responsible disclosure process flagged this capability first, or whether the government identified it independently.
- How "foreign national" gets operationalized. KYC at signup? IP geofencing? Workforce attestations from enterprise customers? Each has very different security and privacy implications.
- Whether allied governments — UK, Canada, Australia, the rest of Five Eyes — quietly get carved out.
Anthropic's usage policy and any forthcoming transparency note are the documents to watch. I'll update as primary sources land.
