The Trump Administration Just Banned a Commercial AI Model for the First Time. We Discuss Was That the Right Call.

Last updated on June 18, 2026

Influential Visions: AI in Government — Part Four of Four

On Friday 12 June, the United States government ordered Anthropic to switch off its two most capable artificial intelligence models for every customer on the planet. It was the first time Washington has used export-control law to disable a commercial AI model, rather than a chip, a weapon, or a piece of hardware.

Steven J. Manning, my co-founder at MONDAY INFLUENCER®, calls the decision a strategic masterstroke rather than a heavy-handed mistake. He closes our four-part Influential Visions series on AI in government with the most contrarian argument of the run.

What happened

Anthropic launched Claude Fable 5 and Claude Mythos 5 on Tuesday 9 June. Fable 5 went to the public, fitted with safety classifiers that decline high-risk requests on cybersecurity, biology, and chemistry, and quietly hand the conversation to the less capable Claude Opus 4.8 instead. Mythos 5, the same model with those classifiers lifted, went only to a vetted group of cyberdefenders and infrastructure operators, through a government collaboration called Project Glasswing.

That programme had already grown fast. It began in April with around fifty partners testing an earlier preview model against their own code. By the start of June, Anthropic had expanded it to roughly two hundred organisations across more than fifteen countries, hunting for the kind of flaw that, left unpatched, could take down a hospital, a power grid, or a bank.

Then, at 5:21pm Eastern on the Friday, a letter from Commerce Secretary Howard Lutnick reached Anthropic chief executive Dario Amodei. Citing national security, the Trump administration ordered the company to block every foreign national, including its own foreign-born staff, from both models, wherever they were in the world.

Anthropic could not check a user’s passport in real time. So at the end of a Friday, while most of its customers were closing their laptops for the weekend, the company’s two newest, most capable models went dark, worldwide, for everyone, with no published date for return.

According to The Wall Street Journal and other outlets, the chain of events began with Amazon. Chief executive Andy Jassy told Treasury Secretary Scott Bessent that Amazon’s own researchers had used a sequence of prompts to coax Fable 5 into producing information useful for a cyberattack. Amazon happens to be one of Anthropic’s largest investors, with a reported thirteen-billion-dollar stake and a hundred-billion-dollar cloud-spending commitment running the other way.

Anthropic disputed the scale of the problem without disputing the finding. The company said the technique surfaced only a handful of minor, previously known weaknesses, reproducible on other publicly available models including OpenAI’s GPT-5.5, and warned that applying this standard across the industry would halt frontier AI releases altogether. The timing could hardly be worse: Anthropic had filed confidentially for a public listing only weeks before.

Two days after the shutdown, more than eighty cybersecurity executives, from firms including Nvidia, Adobe, Sophos, and Zoom, signed an open letter to Lutnick and National Cyber Director Sean Cairncross asking for the directive to be lifted. As I write this, the ban is still in place, and Anthropic’s technical team is in Washington trying to negotiate it away.

Manning’s case

Manning’s opening question sets the whole episode in tension. Is this national security, he asks, or the most expensive competitive takedown in the history of technology?

He starts with history. Fire, the wheel, electricity, the internet, every transformational technology arrived with predictions of ruin that did not come true. He points to Walmart’s plan to retrain, rather than replace, two and a quarter million staff as evidence that the simple story of AI taking jobs misses what is actually happening on the ground.

He casts Dario Amodei less as a creator who lost control of his own invention, and more as a Promethean figure: the one person in the room loudly warning regulators and banks about the risks of what he built, including flagging directly to the Federal Reserve that the model could be used to breach them.

Two numbers keep him up at night, pulling in opposite directions. The United States has poured roughly twenty-three times more money into AI this year than China has, and yet China sits only three per cent behind on capability. Separately, AI models good enough to be genuinely useful are now running locally, with no frontier provider required at all. Either way, in his words, the cat is out of the bag.

He reaches back into his own career for the precedent that matters most to him. In the 1980s, six fierce competitors in direct-response marketing faced a privacy backlash and chose, within a single day, to regulate themselves rather than wait for government to do it for them. His argument: an enforcement budget never runs out and never goes away, but a competitor can, and every AI lab now faces that same choice.

He does not dismiss the underlying worry. His own illustration, of one technically fluent individual working alone, doing real financial damage to a bank without an enterprise’s worth of engineers behind them, makes the point that capability in the wrong hands does not need scale to be dangerous. What he disputes is the government’s method, not the underlying concern.

As a naturalised American, Manning says plainly that he is comfortable with the United States taking a unilateral advantage here, even at a short-term cost to Anthropic’s shareholders and to Amazon’s own position. The bigger constraint on the AI race, he argues, citing Sam Altman’s own warning about OpenAI running out of money within two years, may not be policy at all. It may be the unglamorous business of building power plants fast enough to keep the data centres running.

China’s own economic position, not American export rules, is what Manning believes will actually slow Beijing down, citing long-time China-watcher Gordon Chang. He closes with the AI use cases that get almost no attention: AI shaping planning decisions in New York, AI guiding precision tools inside orthopaedic surgery. Adoption, in his telling, is already further ahead, and already proving its worth, than the policy fight in Washington suggests.

Why it matters

For anyone running a business, the lesson is not really about Anthropic, Amazon, or any single administration. It is that a vendor you depend on can disappear from your stack on a government’s timetable, with no warning and no published return date. That happened to enterprise customers, banks, and government agencies running Mythos-class models for serious work, on a single Friday evening.

The leaders who weathered that weekend without disruption were the ones who had already built optionality into their tools, rather than betting an entire workflow on whichever model happened to be strongest that week. That discipline, not a stronger opinion about whose AI policy is right, is the real competitive advantage on display here.

Four episodes, one question

This closes our four-part series on AI in government. Nicolas Babin opened from inside the European Union. Tushar Kansal reported from India. Darrell Mann carried the question to the United Kingdom. Manning finishes it from inside the country that, for now, still writes the rules everyone else reacts to.

Different place, same question every time: who controls the model, and who decides what it is allowed to do? Across all four conversations, the same shift kept surfacing, from an era defined by general-purpose foundation models built by a handful of labs, to a Sovereign AI Era (countries and companies insisting on direct control over which AI capability operates inside their own borders, rather than renting it from somewhere else).

That shift is not a reason for alarm. It is a sign that AI has become important enough for governments, investors, and security teams to fight over in public, in real time, which is precisely what happens to every technology on its way from novel to essential. The leaders who treat this moment as signal, not noise, will be the ones still standing once the dust settles.

Subscribe to MONDAY INFLUENCER®: https://mondayinfluencer.com More tools and frameworks: https://natschooler.com/tools