Washington's fight with Anthropic is no longer just about one jailbreak. It's becoming a test of whether the U.S. government can decide which frontier AI models the rest of the world gets to use.
On June 12, Commerce Secretary Howard Lutnick's department told Anthropic to cut off foreign access to Fable 5 and Mythos 5, its two most powerful models. The Wall Street Journal reported that the order covered foreign governments, companies and individuals, including foreign nationals inside the United States. Anthropic couldn't reliably separate those users in real time, so it shut the models off for everyone.
That's the part you should sit with. A rule aimed at foreign users quickly became a global suspension, including for American customers, because cloud AI doesn't behave like a crate of chips stopped at a port. If Washington wants to treat model access like an export, the practical result may be blunt shutdowns, not tidy compliance.
The trigger was a reported jailbreak. The Verge reported that Anthropic said the government believed it had learned of a way to bypass Fable 5's safeguards, while Anthropic described the issue as a "potential narrow, non-universal" vulnerability. Business Insider reported that officials grew alarmed during a rapid series of calls involving senior administration figures and Anthropic CEO Dario Amodei. The administration saw national security risk. Anthropic saw an overbroad reaction to a flaw it said wasn't unique to its systems.
Both things can be true, but they don't carry the same consequence. A model flaw can be patched. A precedent is harder to unwind.
Over 50 cybersecurity leaders and technology executives pushed back in a public letter to Lutnick, according to reports from The Verge and Times of India. Alex Stamos, now chief product officer at Corridor and a former Facebook chief security officer, helped organize the effort. Their point wasn't that Anthropic's models are harmless. It was that taking powerful defensive tools away from researchers and companies while rival models remain available may make defenders weaker, not safer.
That is a serious argument, not industry whining. Fable 5 and Mythos 5 were drawing attention because of their ability to find and work through software vulnerabilities. Those same skills can be abused, but they can also help banks, hospitals, energy companies and software teams patch systems before attackers get there. You don't reduce that risk by pretending the capability disappears when an American model is switched off.
On June 23, Legion LegalTech Corp filed suit over the order, TechRadar reported, citing Reuters. The company said it had contractual access to Fable 5 and lost that access because it employed Canadian developers working remotely from Canada. Its core legal claim is simple: current export controls don't clearly cover access to hosted AI models or their outputs. That's not a minor technicality. If a court agrees, Washington may need Congress or a new rulemaking process before it can keep doing this.
The lawsuit also gives the story something most AI policy fights lack: a harmed customer with a billable product, remote workers and a concrete reason to complain. That's useful. Abstract debates about frontier risk can drift into theater. A customer losing access to a tool it had already integrated is harder to wave away.
OpenAI got the softer version
The OpenAI news on June 26 makes the Anthropic fight look less isolated. Business Insider, The Guardian, Axios and AP all reported that OpenAI limited access to its new GPT-5.6 models after a U.S. government request. The models are called Sol, Terra and Luna, with Sol described as the most powerful version. OpenAI is offering access first to a small group of trusted or approved partners while the government review continues.
OpenAI's public posture was careful. It cooperated, but it also said this kind of government access process shouldn't become the long-term default. That's the right warning. Once model releases start moving through customer-by-customer political approval, the line between safety review and industrial control gets thin very quickly.
Frankly, the administration is already acting as if a pre-clearance regime exists, even while the formal machinery is still being built. AP reported that President Trump signed an executive order calling for a 30-day government vetting process for advanced AI systems. Axios reported that officials expect a classified process for assessing certain frontier models by August. You don't need to guess where this is going. The review process is taking shape around live product launches.
For Anthropic, the immediate question is whether it can restore Fable 5 and Mythos 5 without agreeing to conditions that make the next shutdown easier. For OpenAI, the issue is whether a voluntary preview period becomes a routine gate before any major model reaches ordinary customers. For everyone else, the message is already clear: political risk is now part of frontier AI deployment.
The government has a real problem to solve. Advanced models can help people find software flaws, automate attacks and compress work that used to require scarce expertise. But a rushed export order aimed at one company won't settle that problem. It just tells every lab, customer and foreign government that access to American AI can change with a letter from Washington at 5:21 p.m.
Also read: Regulators are finally building the AI they need to police the markets they oversee • Meta is building a prediction market app called Arena and the existing players should be worried • The New York Times is now arguing Microsoft built a machine specifically designed to steal journalism
