The Off Switch Isn't Yours: What the Fable/Mythos Shutdown Taught Me About AI Dependence

In June 2026, a model that a lot of people were building on simply blinked out. The US government issued an export-control directive ordering Anthropic to cut off all foreign access to its two most capable models, Fable 5 and Mythos 5, on national-security grounds. Anthropic complied in about 90 minutes — and to comply at all, it had to pull both models globally. Every user on Earth, paying enterprise customers and Anthropic’s own employees included. The reporting tied the urgency to a suspicion that a China-linked group had already accessed Mythos, alongside a jailbreak that could coax the model into surfacing software vulnerabilities.

Anthropic disputed the framing — calling the vulnerabilities minor and findable with publicly available models, and arguing that governments may block unsafe deployments but only “as part of a statutory process that is transparent, fair, clear, and grounded in technical facts,” which this wasn’t. You can read that as the government overreaching or as it acting prudently. Either way, the underlying fact doesn’t move: the most powerful version of the tool went dark for everyone, overnight, by an order none of its users had any say in.

That’s the part I keep sitting with. We talk about the AI access gap as a question of money or skill — who can afford the good model, who knows how to drive it. The Fable/Mythos shutdown stripped the abstraction off and showed a third axis underneath: the switch lives somewhere else entirely. It isn’t yours. It belongs to a handful of companies and the governments that can countermand them, and any of them can flip it between one afternoon and the next without asking you.

If you’ve built real work on a hosted model — a workflow, a product, a research habit, a way of thinking through hard problems — that should land. The leverage is genuine, which is exactly what makes the dependence genuine too. A tool you can’t keep running is a tool someone else can take back. That’s not unique to AI; it’s true of any utility you rent instead of own. But the leverage here is large enough, and the recent shutdown abrupt enough, that pretending the off switch is yours is just a comfortable lie.

So here’s how I actually hedge, without giving up the upside:

• Don’t single-thread on one model. I keep more than one provider in reach, and I keep at least one capable open-weights model I can run locally as a floor. It won’t match the frontier, but it can’t be revoked by directive — the worst case is “slower,” not “gone.”
• Own the inputs. My prompts, my context, my notes, the structure of how I work — those live in plain files I control, not locked inside one vendor’s chat history. If a model disappears, the scaffolding I built around it doesn’t.
• Keep the thinking in my hands. This is the deeper hedge. If I can still reason through the problem when the tool is offline — slower, sure — then a shutdown is an inconvenience, not a collapse. The danger isn’t losing the model; it’s having outsourced so much that you can’t function without it.

None of this is a reason to stay away. The tool is too good for that, and I’m still firmly optimistic about it. It’s a reason to hold the optimism the right way: build with the frontier, lean on it hard, and assume it could be pulled out from under you anyway. Plan for the off switch you don’t control, and a shutdown becomes a bad afternoon instead of an existential one.

That’s the eyes-open version. Use the best tool you can, and don’t forget whose hand is on the switch.

I pulled this out of a longer piece on using AI without handing over your judgment — the full argument, costs and upsides and all, is in Optimistic, Eyes Open.