The abrupt shutdown and rapid restoration of Anthropic’s powerful Fable 5 and Mythos 5 AI models has exposed deep tensions over how the U.S. should regulate frontier artificial intelligence while trying to stay ahead of global rivals.
Early releases and rising concerns
In April, Anthropic began offering its Mythos‑class systems to a limited set of organizations, warning that the models were unusually capable at “identifying and exploiting vulnerabilities in software.”1 A more constrained public version, Fable 5, followed in June with additional security guardrails.1
On June 12, the Trump administration moved to clamp down. The Department of Commerce added Fable 5 and Mythos 5 to an export‑restricted list, effectively barring access by foreign nationals — including some Anthropic employees — and forcing the company to “abruptly disable” both models for the public.2 Officials cited fears that adversaries could weaponize the systems’ hacking capabilities.2
Weeks of dispute and negotiation
Anthropic argued the order stemmed from a misunderstanding of a possible Fable 5 “jailbreak” and sent executives to Washington to negotiate with the White House.2 Cybersecurity experts questioned the policy, viewing it less as a security fix and more as leverage over a company whose leaders had criticized how politicians might use advanced AI.1
Meanwhile, U.S. officials faced mounting pressure as Asian competitors like Fugu and Tulonfeng approached Mythos‑level capabilities, raising concerns that American firms could fall behind.1
Restrictions lifted, safeguards tightened
After weeks of talks, Commerce Secretary Howard Lutnick announced that Anthropic had agreed to “proactively detect and address security risks,” work with the government on standards for Mythos, Fable and future models, and report malicious activity.1 On June 30, the Trump administration lifted export controls on Claude Fable 5 and Mythos 5, with access returning to customers starting July 1.3
The government and Anthropic highlighted a new safeguard that blocks the jailbreak officials feared 93% of the time; in the remaining cases, outputs were limited to already known or patched vulnerabilities, though benign queries may now be over‑blocked.4
Competing visions of AI governance
The administration framed the outcome as a model of public‑private collaboration to “strengthen America’s leadership in AI,”4 while Anthropic cast it as a hard‑won resolution that lets it restore service and continue toward an anticipated IPO.2 Yet with OpenAI’s latest systems and Mythos 5 still funneled to select government‑approved customers, the episode underscores how U.S. AI policy is evolving through ad‑hoc deals rather than clear, durable rules.4