The Switch Someone Else Holds

The modern digital economy runs on a persistent illusion of stability. Organisations build intricate, automated workflows on the assumption that the tools they depend on will stay available, governed by ordinary market forces. On 12 June 2026, at 5:21 PM Eastern Time, I watched that illusion get dismantled. I did not watch it as a neutral observer. I build on the opposite principle, distributed, minimum-dependency architecture, so I recognised the cutoff at once: it was the failure mode I had spent my work trying to design out.

This was not a voluntary safety pause or a scheduled maintenance window. It was a federal export-control directive from the United States government. By targeting "foreign nationals", a categorisation so broad it barred even Anthropic's own international employees, the directive forced an immediate global shutdown of the Claude Fable 5 and Mythos 5 models. The trigger was an alleged method of bypassing the safeguards in Fable 5. Anthropic disagreed with the proportionality of the response, arguing the demonstrated vulnerabilities were narrow, non-universal, and similar to flaws already discoverable in other publicly available models. The legal reality held regardless: because the directive covered all foreign nationals worldwide, disabling the models entirely was the only way to comply.

A crowded, partisan debate ignited at once across European policy circles. Many treated the cutoff as a sovereignty outrage, a geopolitical weapon aimed at foreign innovation. Yet experts in international and data protection law have pointed out that the anger may be running ahead of the facts: the foreign-only scope was likely an artefact of the only legal tool available for a rapid freeze, not a targeted strike against allies.

But the motive does not matter. Whether the directive was a deliberate weapon or the blunt application of a domestic security law, the architectural result was identical. Workflows halted. The mechanism existed, and it was used. The fact that the switch can be pulled, for any reason, is what reveals the structural flaw.

When an enterprise adopts an advanced capability, it reorganises around it. The engineers stop performing the core task and start managing the model's outputs. When the model is switched off by a single directive, the enterprise does not just lose a tool; it loses operational capacity, because the human architecture has already been dismantled to accommodate the machine. That is the danger of a switch someone else holds.

The sudden cutoff is the acute version of a much slower decay, the pattern the writer Cory Doctorow calls "enshittification". It runs in three phases. First, the bait: a platform floods users with value, free services, steep discounts, directing all its surplus toward winning a base. Then lock-in: once a critical mass has settled in, the platform makes leaving prohibitive through proprietary formats, cultivated habits, and deep integration into business workflows, and pivots to court suppliers with favourable terms before locking them in too. Finally, extraction: with users and suppliers unable to migrate without heavy cost, the platform degrades its own service to claw value back for shareholders, replacing organic results with paid placements and restricting access.

Doctorow names a second pattern that compounds this: the "reverse centaur". A centaur is a human augmented by a machine. A reverse centaur is a human reorganised to serve the machine, reduced to an appendage of a workflow it no longer controls, often there only to absorb blame when the automated system fails. Either way, the business no longer controls its own destiny.

The export directive and slow platform decay are the same dynamic at different speeds. Degradation extracts value through a slow tightening of terms; the kill switch extracts immediate compliance. Both draw their power from one structural flaw: the concentration of reliance. When a technology becomes critical infrastructure for millions of businesses, control over access becomes a lever, and the leverage someone holds is exactly proportional to how much you depend on them. The defence against both the slow squeeze and the sudden shock is the same. A workflow distributed across several providers, leaning only lightly on any single proprietary node, cannot be quietly degraded for profit or neutralised by one directive.

Before the directive proved the fragility of foreign dependencies, strategic foresight researchers had already warned of it. The Europe 2031 report, by a coalition of European technology and policy experts (Daan Juijn, Stan van Baarsen, Judith Dada, Lily Stelling, Philip Fox, Alex Petropoulos, and Michiel Bakker, with editing by Tom Chivers) projected Europe's slide into irrelevance through dependency on foreign-controlled chokepoints, dramatised through the eyes of a policy worker and climaxing in a Washington negotiation over control of ASML, the sole maker of the lithography equipment needed to print cutting-edge chips. The report published essentially as the export directive hit, making its warning real faster than its authors imagined. Renting access to a frontier model gives the renter no power to keep that model from being disabled by a single regulatory stroke. The dependency was always the vulnerability.

This forces a re-evaluation of what sovereignty means. Policy debate has been overwhelmingly geographic, fixated on where the servers physically sit. But the location of a server rack offers little protection if the software on it is controlled by an entity subject to sudden legal injunctions. A border does not stop a kill switch. And the opposite reflex, demanding isolation on the argument that regions are "better off without" foreign models, is one-sided disarmament: it is not sovereignty, it is forced technological regression.

Real control is architectural, not geographic. Digital sovereignty is the operational ability to understand, secure, modify, and replace your own systems, and to govern what you depend on. If a critical workflow is anchored to a single proprietary model, disabling that model collapses the workflow. But if the workflow is built on interoperability, with an abstraction layer that routes requests across providers by real-time availability, a single directive can turn off a specific model and still not turn off the system. The same logic answers enshittification: keep the functional ability to exit a degraded platform without destroying the business.

The constructive model is not in the big tech hubs but in Estonia. One of the most digitised states in the world, its e-governance rests on traditional cryptography that a sufficiently capable quantum computer could break, exposing decades of confidential state data through "harvest now, decrypt later" attacks. Rather than wait for foreign giants to patch this, Estonia treated post-quantum resilience as an immediate national-security matter and contracted Cybernetica, a domestic firm in Tallinn and Tartu, to lead the transition.

Cybernetica is a study in architectural sovereignty. It keeps a slight majority of its staff, 51 percent, in the university town of Tartu, cultivating in-house mathematical depth. As its research director Dan Bogdanov notes, Estonia is unusual in how much of its core technology it builds from the ground up rather than consuming from hyperscalers; a German professor once called it the land of "exotic, gourmet cryptography". The work spans a national post-quantum roadmap, the quantum-proofing of the Population Register, and ongoing research to keep the underlying mathematics native rather than outsourced. The result is a state that can audit, modify, and replace its own cryptographic foundations without waiting for permission or an API update from a foreign provider. That is what it means to govern what you depend on.

The lever may be pulled slowly, over years, by a board chasing higher revenues, or instantly, by a security apparatus enforcing a mandate. The result is the same: the capacity to operate independently is lost. The durable defence is disciplined subtraction. We reduce the leverage someone holds by reducing what we depend on, building on interoperable standards, distributing across interchangeable providers, and keeping a real exit. Sovereignty proves itself by needing less, not by isolating more.

And note what this analysis has quietly assumed: that the cost of a pulled switch is measured in halted workflows and lost productivity. For an enterprise, it is. But the same structural exposure, applied to people rather than companies, is measured in something far graver and far less evenly distributed. That is a subject for its own reckoning. For now it is enough to see the architecture clearly: the switch, whoever holds it and for whatever reason, has power only over what we have made ourselves unable to live without.