Last week, frontier model releases were the biggest national-security story. This week, the next phase is clear, access is being permissioned by the government. The U.S. lifted export controls on Anthropic's Fable 5 and Mythos, but they are still not open access across the world. Fable 5 came back globally with additional safeguards, but Mythos 5 remained limited to approved organizations. Anthropic also committed to deeper government collaboration, pre-release testing, and shared jailbreak standards. The biggest questions for releasing models are who will get access, under what control, and after who reviewed it.
AP gives the cleanest update to last week's shutdown story: the Trump administration lifted restrictions on Anthropic's newest Claude models after a cybersecurity alarm, but access did not simply return to normal. Fable 5 is widely available again, while Mythos 5 is being restored only to a select group of U.S.-based organizations approved by the federal government. That makes this week's real story permissioned access: frontier AI can be paused, reviewed, restored, and segmented by customer type.
Reuters is the best business-facing version of the access-restoration story. It emphasizes that the Commerce Department lifted the curbs less than three weeks after imposing them, after Anthropic implemented safeguards and began working with the government and Glasswing partners on shared standards for jailbreak assessment. The important business point is that model access is becoming an operational dependency: customers, cloud partners, and AI labs now have to plan around government review as part of frontier-model deployment.
Anthropic's own post is the primary source for what changed after the restriction was lifted. The company says Fable 5 is returning globally, Mythos 5 access has been restored for a set of U.S. organizations, and broader Glasswing access is still being coordinated with the government. The post also describes the safeguard update: an improved safety classifier can block requests matching the reported bypass and route them to a less capable model. This is permissioned AI in practice: access tiers, model fallbacks, government coordination, and shared safety standards.
Lawfare provides the strongest strategic-policy read for this issue. The piece argues that using export controls for AI models deserves a fair hearing when a system could enable serious offensive cyber capability, but it also warns that chip-style controls may not translate cleanly to model access. This helps separate this week from last week's release drama. The deeper issue is whether the U.S. can create a durable permissioning regime without pushing allies, researchers, customers, and competitors toward alternatives outside U.S. control.
The Guardian gives the global-access angle: Fable and Mythos were restricted because U.S. officials feared the models could be misused for serious cyberattacks, then restored after Anthropic added safeguards and committed to closer government cooperation. The piece also captures the controversy around customer vetting, including concerns that government-approved access lists could shape who benefits from the best models. This is the reader-friendly version of the week's core tension: safety review may be useful, but permissioned access also concentrates power.
#global-access#customer-vetting#ai-policy
Going Deeper
Optional reads for those who want more. (Some may be behind a paywall)
A Kill Switch for Frontier AILawfareHelpful legal context on how export-control authority turned the Mythos/Fable order into a practical global shutdown, and why model access may become harder to separate from national-security law.
A Red-Team Study of Anthropic Fable 5 & Opus 4.8 ModelsarXivRecent research for readers who want the technical backdrop: frontier-model safeguards can be strong in aggregate while still leaving meaningful residual risk under sustained automated jailbreak attempts.