A few weeks ago, Anthropic launched its most powerful model yet. Three days later, it was gone. Not an outage. Not a bug. The US government ordered it shut off, worldwide, for every customer — including the people paying for it.
If you’d built anything on that model in those three days, you found out the hard way that there’s a risk category most of us weren’t pricing in: the government deciding you don’t get access to the best AI tool anymore, with no warning and no explanation.
What actually happened
Anthropic released Claude Fable on June 9th. It’s a guardrailed version of their most capable model, Mythos, which they’d kept locked up because they said it was too good at finding security vulnerabilities to release freely. Fable was supposed to be the safer, public version.
On June 12th, the government stepped in. Some combination of internal testing and outside groups found a way around Fable’s safeguards, which made it look a lot like the more dangerous Mythos model underneath. The government called it a national security risk for non-US use and ordered access cut off. Anthropic couldn’t tell US users from foreign users in real time, so the only option was to turn the whole thing off — for everyone.
Two weeks later, on June 26th, part of it came back — but only for a short list of pre-approved organizations. The rest of us are still locked out. And in the same window, OpenAI’s newest model launched the same way: gated, government-approved access only, before the general public ever saw it.
How we got here
This didn’t come out of nowhere. Back in February 2026, Anthropic and the Pentagon got into a dispute over contract terms — the Pentagon wanted language that would let Claude be used for mass surveillance and autonomous battlefield decisions, and Anthropic said no. That led to a federal contract blacklist in March and active litigation that’s still going. Then on June 2nd, a new executive order asked AI companies to voluntarily share new models with the government for a 30-day review before release.
Voluntary, on paper. In practice, the government used existing export control authority to force the issue ten days later. There’s no public benchmark, no clear legal standard, and no real transparency into what specifically triggered the shutdown. Just a directive, and a model that disappeared.
Why this matters for your firm
If you built a workflow, a client deliverable, or an internal tool that only works with one specific model, that thing can become useless overnight — and not because of anything you did. There are already lawsuits against Anthropic from people who upgraded to use Fable right before it vanished.
There’s also a quieter risk: indirect exposure. Any tool you’ve licensed that calls a frontier model under the hood inherits that model’s regulatory risk, even if you never typed a prompt into it directly. And if you’re at a larger firm, there’s a competitive angle too — if only one of the Big 4 gets approved access to the next gated model and the others don’t, that’s a real edge, handed out by a government list instead of earned in the market.
Key Takeaways
- Regulatory shutdown is a new vendor risk category. It’s separate from uptime, security incidents, and pricing — and no current SLA covers it.
- “Foreign national” restrictions can take out your whole org. Providers often can’t segment access by nationality in real time, so everyone loses access together.
- Audit indirect exposure. Know which underlying model powers every AI tool you’ve licensed, not just the brand name on the invoice.
- Add a question to vendor due diligence: what’s your exposure to export-control action, and what’s the contingency plan if a model gets pulled?
- Stay model-agnostic where you can. Diversifying across providers and models is no longer just a cost play — it’s a continuity plan.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: AI Government Shutdown Risk
Leave a Reply