On June 12, 2026, Anthropic announced that the US government had ordered it to suspend all access to two of its most capable models, Fable 5 and Mythos 5 immediately, worldwide, for every user. Not a price change. Not an outage. Not a deprecation with six months' notice. A regulator decided, and the most advanced models from one of the best-funded AI labs in the world went dark for everyone.

The stated reason was a national-security concern: authorities flagged a "narrow" way to get the model to perform cybersecurity tasks it wasn't supposed to. Anthropic disagrees with the decision and is working to restore access. Whether they're right isn't our point. Our point is what this episode reveals for anyone running a business on top of AI.

Your most powerful model is also your most fragile dependency

The instinct is to reach for the newest, most capable model available. It's also, almost by definition, the one most likely to draw regulatory attention to get rate-limited, repriced, or, as we just saw, switched off by someone other than the vendor. The frontier is exciting. It is not stable ground to build your daily operations on.

If one specific model is wired into how your team works quoting, drafting, triaging tickets, answering customers then the day that model disappears, so does that part of your business. You've inherited a single point of failure you don't control and can't appeal.

Build so you can swap

The protection isn't to avoid AI. It's to avoid depending on one irreplaceable piece of it. The practical version for an SMB:

  • Keep your data, your prompts and your business logic separate from any one provider, so that moving to another model is a configuration change, not a rebuild.
  • Prefer the simplest model that does the job well over the most powerful one that's overkill. Less capable often means more stable, cheaper, and less likely to vanish.
  • Know your fallback before you need it. If your model went away tomorrow, what's the plan? "We'd be stuck" is the wrong answer.

The sovereignty angle Quebec businesses can't ignore

There's a second lesson here, and it's a familiar one. The switch was flipped by a foreign government, reaching through a foreign vendor, into the tools of users everywhere here included. We usually talk about data sovereignty in terms of where your data sits and who can read it (Law 25). This adds a dimension: who can turn off the capability you depend on.

You can't regulate Washington. But you can make choices about which models, which vendors, and how much of your operation rides on infrastructure governed by rules made elsewhere. For some workloads, a model you can run under Canadian jurisdiction isn't just a compliance checkbox. It's continuity insurance.

None of this means "wait"

The lesson is not to fear AI or sit it out the businesses that adopt it are pulling ahead. The lesson is to adopt it like a grown-up: start from the problem, choose the simplest tool that solves it, and build so that no single vendor, model, or government holds the off switch to your operations.

That's exactly what a diagnostic is for. Before any build, we map what you actually need, how much capability the job truly requires (often less than you'd think), and where your dependencies and risks really lie. The goal isn't the most impressive model. It's a solution that still works on Monday no matter what happened over the weekend.

If you're not sure how exposed your AI plans are, ask for a diagnostic. No obligation, just a clear picture of where you stand.