Last week we wrote that a model you build your business on can be turned off by someone who is not you. That was the abstract version. The reporting since has filled in the blank, and the answer is uncomfortable. It was not only the government. One of Anthropic's largest investors reportedly helped push the model offline.

On June 12, 2026, the US ordered Anthropic to suspend its Fable 5 and Mythos 5 models for any foreign national, inside or outside the country, citing national security. Anthropic could not verify nationality in real time, so it shut both models off for every customer in the world. Other models, including Claude Opus 4.8, stayed up. We covered the mechanics of that shutdown in our piece on the Fable 5 government order.

The continuity risk we flagged in the abstract now has a mechanism. A rival can help lobby your model offline.

The new fact

The order did not come out of nowhere. According to The Verge, reporting the Wall Street Journal, the export-control directive was triggered in part by cybersecurity research from Amazon and by conversations between Amazon CEO Andy Jassy and the White House.

Read that again. The research that helped get the model banned came from a company that has put billions into the company that makes it. Amazon is one of Anthropic's biggest backers. It also runs the cloud a lot of Anthropic's compute sits on. And it has its own frontier-model ambitions.

So the entity that supplied the security case against Fable was not a neutral lab or a government red team. It was a shareholder with a competing product.

The investor conflict

This is the part that should bother anyone building on a frontier model. Jassy and executives from five other companies reportedly warned the Trump administration about security vulnerabilities in Fable, according to The Decoder. The same report notes the obvious tension. Amazon is one of Anthropic's largest investors.

You can hold both things in your head at once. Maybe the vulnerabilities are real and worth flagging. Maybe a company that competes in the same market and funds its rival is exactly the wrong party to be carrying that flag to the government. Both can be true, and that is the problem. The motive is impossible to clean.

The damage here is structural. A backer that also sells a competing model has a reason to want Fable slowed down, and now we know one had the ear of the people who can act on it. The vulnerabilities may be genuine. The party that carried them to Washington still had a motive with nothing to do with safety, and there is no way to separate the two from the outside.

That is the conflict. It is not hypothetical anymore.

The national-security thread

There is a second strand, and it is not about competition. Part of the decision was driven by fears that a group linked to China had accessed Mythos, according to The Verge, reporting Semafor. That is a genuine security concern, and it explains why the order reached for export-control machinery rather than a normal product complaint.

The stated worry is a familiar one for anyone who has watched the jailbreak debate. A powerful model in the wrong hands becomes a tool, and the people raising the alarm framed Fable and Mythos as exactly that kind of risk.

Here is where the two strands tangle. A real foreign-access fear and a self-interested competitor's research arrived at the same conclusion at roughly the same time. When the security case and the commercial case point the same way, it gets very hard to tell which one is actually steering. We cannot tell from the outside. Neither, probably, can most people inside.

The backlash

The security experts are not sold. Dozens of cybersecurity veterans urged the White House to remove the restrictions, arguing the order limits defenders' ability to secure software, according to TechCrunch. Their case is that the most capable models are tools for the people defending systems, not just the people attacking them. Take the best model away from defenders and you do not make anyone safer. You make the defense weaker.

The government side is unhappy too, for a different reason. Officials accuse Anthropic of releasing Fable 5 without approval, according to The Decoder. One administration official put it bluntly: "They screwed us." Talks are now underway with the Department of Commerce and the CIA.

For its part, Anthropic says it reviewed a demonstration and found a small number of previously known, minor vulnerabilities. It disagrees with the decision. It is complying anyway while it works to get access back. That is the posture of a company that thinks it is right and also cannot afford to fight.

What it means

Last week the lesson was a risk: a third party can switch off your model. This week the risk has a named cause. The third party can be a rival, and the rival can use security politics to do it.

If you run anything on a frontier model, that changes your planning. The model is not a stable input you have a contract for. It is a thing that can be pulled by a government acting partly on research from a company that competes with your supplier. There is no service-level agreement that covers that.

The defense is the boring one we keep coming back to. Build so the model can be swapped. Abstract the model behind your own interface. Keep a second provider warm. Know, before you need to know, what breaks if Opus 4.8 is the only thing you can reach by Monday. European banks running on borrowed AI infrastructure face a sharper version of this, which we covered in our piece on sovereign AI and bank dependency.

The companies that came through last week's shutdown with the least pain were the ones that already treated their model as replaceable. That was a design choice made months earlier, not a reaction. Make it now, while it is cheap.

What to watch

The Commerce and CIA talks are the near-term signal. If Anthropic gets Fable and Mythos restored quickly, the episode reads as a negotiation that got loud. If access stays dark for weeks, it reads as something more permanent, and the export-control framing starts to look like the new normal for frontier models.

Watch whether other investors with competing products take the same route. Amazon went to the White House with security research about a model it helps fund. If that works, it becomes a playbook. The next time a backer wants a rival's model slowed down, it now has a template and a precedent.

And watch the standard the government sets. Officials are reportedly pushing Anthropic toward models that cannot be jailbroken, which security researchers say is closer to impossible than hard. A rule no model can meet is a rule that can be used against any model, whenever someone with the right phone number decides to use it.

Sources

If a company can fund a frontier model and help lobby it offline in the same quarter, what does any build-on-it strategy actually rest on?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.