The U.S. just proved it can flip a “kill switch” on powerful AI models—and then quietly turn it back on—without ever showing the public the technical evidence behind its claim.
Story Snapshot
- The Commerce Department ordered Anthropic to globally shut down its Fable 5 and Mythos 5 AI models over a claimed cybersecurity risk, then lifted the export controls less than three weeks later.
- The order used export control law in a new way, treating who can access an AI system as an “export,” even when the model stays on U.S. servers.
- Anthropic says the government never shared detailed, written proof of a serious jailbreak and that the reported flaws were minor and already present in other public models.
- Both right and left see the episode as fresh evidence that a secretive federal bureaucracy can shut down key technology for everyone based on vague claims and unclear legal authority.
How a Little‑Known Export Rule Shut Down Frontier AI Worldwide
On June 12, 2026, the U.S. Department of Commerce ordered Anthropic to cut off its two most advanced AI models, Fable 5 and Mythos 5, to every foreign national, inside or outside the United States. Because Anthropic could not reliably sort users by citizenship, it shut the models down for everyone, including U.S. users and its own staff. For 18 days, the most capable American AI system was simply gone, not by court order or public law, but by a single export-control directive.
The order relied on powers under the Export Control Reform Act and existing export rules that usually govern physical goods, chips, and software shipped overseas. Legal analysts say this was the first time Washington treated remote AI access itself as an export, stretching old rules into a new digital space with almost no public debate or clear limits. In practice, it showed that a small group of officials could treat advanced AI as strategic infrastructure and pull the plug worldwide when they decide a risk exists.
Competing Stories About the Alleged AI “Jailbreak”
Officials say the move came after a “highly credible, trusted partner” showed a way to bypass Fable 5’s safety guardrails and use the model to find software weaknesses that could help foreign cyberattacks. Reports point to Amazon, a key Anthropic investor and cloud provider, as the company that raised the alarm after internal tests. From the government’s view, stopping hostile intelligence services in China or Russia from using U.S. AI as a hacking tool justified fast, tough action.
Anthropic tells a very different story. The company says the government never shared detailed, written technical evidence, only “verbal” claims about a narrow, non‑universal jailbreak. After reviewing the same test, Anthropic concluded the flaws were minor, already known, and similar to issues that other frontier models like GPT‑5.5 can also find. The company stressed that no universal jailbreak had ever been found in thousands of hours of joint red‑team testing with U.S. and U.K. government experts. To Anthropic, the episode looks less like stopping a clear and present danger and more like a rushed reaction to a narrow bug.
A New AI “Kill Switch” and Old Fears of Government Overreach
When Commerce Secretary Howard Lutnick warned of “swift criminal and civil penalties” for non‑compliance, the message was clear: obey or face serious legal trouble. Because the directive covered both model weights and certain “dangerous outputs,” the government effectively claimed power to decide not just where the AI runs, but what kinds of answers it can lawfully give to foreign users. Commentators at Lawfare and other outlets have described this as a de facto “global kill switch” for frontier AI, set off by a small, secretive group of officials.
fable 5 came back from the US export ban today.
anthropic also shipped claude science + sonnet 5 today.anthropic in Q2 2026:
• fable 5 (frontier model)
• mythos 5 (gated cyber model)
• opus 4.7 + 4.8
• claude cowork GA
• claude code (huge updates)
• claude design
•… https://t.co/iI3zmR8ln5— capGoblin (@dharshan_tw) July 1, 2026
For many Americans, this hits a nerve that crosses party lines. Conservatives who already distrust the “deep state” see an unelected bureaucracy using vague national‑security claims to override markets, contracts, and even access for U.S. citizens. Liberals who worry about corporate power and civil rights see a new tool that could easily be turned on activists, researchers, or disfavored firms, with no transparent process or court review. Both sides share one core concern: the rules are unclear, the evidence is hidden, and the people making these decisions do not answer directly to voters.
Why the Rapid Reversal Raises More Questions Than It Answers
Less than three weeks after the shutdown, the Commerce Department lifted the export controls and allowed Anthropic to bring Fable 5 and Mythos 5 back online. Anthropic agreed to work closely with the government on new testing, monitoring, and release standards for these and future models. Yet there is still no public, detailed explanation of what changed. Officials have not released the full letter, the technical report on the jailbreak, or any independent audit showing the risk is now solved.
That quick reversal deepens a sense that advanced AI is now governed by improvised rules. Export‑control lawyers note that traditional regulations were built for hardware and software shipped abroad, not for cloud services that people tap from anywhere. Critics warn that once Washington claims it can decide, model by model, who is “trusted” enough to use powerful AI, every future launch could depend on pleasing distant regulators and their favored partners. In a country already divided over “woke tech,” “America First,” and elite capture, the Anthropic case looks like one more example of a government that can act fast in secret, but struggles to act openly and fairly.
Sources:
theamericanconservative.com, reuters.com, facebook.com, lawfaremedia.org, techpolicy.press, blog.volkovlaw.com, anthropic.com
