Our Take
The government invoked export controls on software that cannot be exported, revealing it has no clear legal framework for AI and is improvising enforcement on the fly.
Why it matters
Anthropic is not the first target and will not be the last. Frontier labs need to know the rules before launch, not after a shutdown order. Without clarity, U.S. AI companies face unpredictable intervention that Europe and Asia will use as proof they should build their own.
Do this week
Legal/Compliance: Document your current safeguard stack and jailbreak mitigation before the next administration action arrives, so you can respond to unannounced government demands with evidence rather than improvisation.
Anthropic shut down Fable 5 and Mythos 5 after an unexplained government order
The Trump administration directed Anthropic to cut access to its newest models for all foreign nationals and overseas users. Anthropic complied, taking both systems offline. The government cited "national security authorities" and an "export control directive" but has not published the legal basis, the specific violation, or the remediation path (per The Verge reporting).
Anthropic stated that the government cited potential jailbreak activity by groups linked to China. The company also clarified that such vulnerabilities did not bypass all of its safeguards, suggesting the order may have been overly broad.
The Trump administration has not explained why export controls, a mechanism designed for tangible goods and discrete data transfers, apply to a cloud-hosted service where users access model outputs but not weights, source code, or copies of the model itself.
Export control law was never written for API access
Hanna Dohmen, a senior research analyst at Georgetown's Center for Security and Emerging Technology, said this is "the first time U.S. export controls have been used to control access to an AI model in this way" (per The Verge). Andrew Reddie, a UC Berkeley professor on export policy, called the regulatory landscape "an open question" and noted the government's lack of clarity on what it expects from model developers.
The Anthropic case exposes three separate governance gaps:
- Export controls historically govern physical goods and downloadable software; remote API access is a known gap Congress is already attempting to close through pending legislation.
- The government has not stated whether Anthropic was targeted because Mythos and Fable are uniquely powerful, because specific safeguards were insufficient, or for political reasons.
- If the order stands, each frontier lab (OpenAI, Google, Meta, xAI) will be forced to guess which capabilities or safeguard thresholds will trigger similar action.
Reddie warned that if the U.S. government sets "impossible to jailbreak" as the de facto standard, it will eliminate American models from competition. Simultaneously, arbitrary intervention without public explanation will push other countries to avoid dependence on U.S. systems entirely.
Prepare for ad hoc enforcement
The incident reflects a contradiction: the Trump administration claims to favor a hands-off approach and American technological leadership, yet it deployed an unexplained shutdown order against a domestic champion. This is not sustainable governance, but it is the reality. Model developers should assume that safeguard design, jailbreak resistance, and user access policies will be contested after launch, not before. Document your threat model, your mitigation strategy, and your audit findings now. If a shutdown order arrives, evidence of good-faith safety effort is your only defense against indefinite offline status. Lobby for public guidance before your next release, but assume it will not come.