# Your AI has an off switch. Who holds it? > A June 2026 AI export control letter switched off two frontier models worldwide. Critical AI needs a path that survives someone else's policy change. - URL: https://www.sector88.co/blog/ai-model-access-export-controls-off-switch - Author: Taskin Karlar - Published: 2026-10-15 - Updated: 2026-10-15 - Category: Analysis - Tags: ai-policy, export-controls, sovereign-ai, open-weight-models ## Quick answer For any hosted AI model, the off switch sits with the provider and with every government whose law reaches it. On 12 June 2026 a US export control letter led Anthropic to switch off two frontier models for every user worldwide for nearly three weeks. Critical AI needs a fallback that runs under the operator's command. On 12 June 2026, the US Department of Commerce sent Anthropic an export control letter requiring a licence before any foreign national, inside or outside the United States, could access its two most capable models, Fable 5 and Mythos 5. Anthropic [said](https://www.anthropic.com/news/fable-mythos-access) it had no reliable way to check nationality in real time, so it switched both models off for every customer worldwide until the controls were [lifted on 30 June](https://www.anthropic.com/news/redeploying-fable-5). As of 15 October 2026, the legal questions are still open. On 6 October, [The Regulatory Review](https://www.theregreview.org/2026/10/06/oktenli-when-a-chatbot-becomes-an-export/) argued that applying export controls to hosted AI models calls for public rulemaking rather than undisclosed directives. That debate belongs to lawyers and lawmakers. The operational lesson belongs to anyone running AI inside a critical system. ## Why AI export controls now matter to operators This was not an outage or a cyberattack, and nothing broke. A lawful letter arrived at 5:21pm Eastern time on a Friday and took effect immediately. Within hours, two frontier models went dark for American users and allied institutions alike. The letter named one company, but its effect was global. Reuters described the action as the first of its kind, and Korea's [Futures](https://futures.re.kr/en/articles/cm-349/) noted that a tool built for semiconductors had been aimed at an API service. [CSIS](https://www.csis.org/analysis/department-commerce-restricted-access-anthropics-latest-models-what-comes-next) wrote that every company must now consider such controls being used again. It expects uncertainty over durable access to push foreign customers toward options they see as more reliable, including small open-weight models they operate themselves. Access is still being settled one decision at a time. On 12 October, the UK's AI minister confirmed that Britain's AI Security Institute would keep early access to the most powerful US models, [AFP reported](https://www.afp.com/en/uk-retains-early-access-test-latest-us-ai-models), following reports that the White House had restricted it. **An AI off switch is any point outside the operator's control where a single decision can stop a model from serving: a licence condition, a government letter, an account flag or a contract term.** Every hosted model has several, and most stay invisible until someone uses one. ## Whose law reaches your inference path? The useful question is not whether your AI provider is trustworthy. Anthropic complied with a legal directive, kept its other models running and restored access within three weeks. The useful question is whose law can reach the path between your data and your model's answer, and how quickly. For a chatbot drafting emails, a three-week gap is an inconvenience. For AI that triages alarms in a grid control room, screens satellite telemetry or supports a defence logistics cell, it is a gap in operations. It arrives with hours of notice, through a channel the operator never sees. We made the same point about physical failures in [the cloud has a blast radius](/blog/cloud-outages-ai-blast-radius): some decisions reach every region at once. ## What it means for operators Hosted frontier models will remain the right tool for many jobs. Critical workloads need a second path that keeps working when someone else's policy changes. In practice: 1. **Map your off switches.** For each AI workload, list who can stop it: the model provider, the cloud, the identity provider, and every government with jurisdiction over them. 2. **Qualify an open-weight fallback.** Choose a model whose weights you hold, test it on your own workload, and know what quality you trade away. 3. **Run it under the operator's command.** Keep the fallback on infrastructure you control, at the site, so it serves when the hosted path is gone. 4. **Keep the record locally.** Log which model and version ran, on what input, and when, so you can show what happened during the switch-over. 5. **Rehearse the switch.** A fallback nobody has tested is a plan, not a capability. This is the "who can stop it" question from our [Australian AI sovereignty submission](/blog/location-is-not-control-australias-ai-sovereignty-test), applied to the AI supply chain. Location is not control, and neither is a contract. For the wider frame, see [what sovereign AI actually means](/blog/what-is-sovereign-ai). ## Where Sector88 fits Sector88 is one system to deploy, run and govern AI at every site, connected or completely cut off. Operators choose which models run, keep a durable record of what ran, and keep serving when the link to a hosted model is gone. See how it works on the [Sector88 platform](/platform). The most important off switch in your AI stack might be a letter you will never see. ## Frequently asked questions ### Can the US government cut off access to an AI model? Yes, and it already has. On 12 June 2026 the US Commerce Department told Anthropic that foreign nationals, inside or outside the United States, needed a licence to access Fable 5 and Mythos 5. Anthropic could not check nationality in real time, so it disabled both models for all users until the controls were lifted on 30 June 2026. ### Is remote access to an AI model an export? It is legally unsettled. Under US export rules an export usually means shipping or releasing software or technology. Users of a hosted model never receive its weights or code. CSIS noted that Commerce had previously treated remote access as outside those rules, and The Regulatory Review argued in October 2026 that the question needs public rulemaking. ### How do organisations reduce AI supplier risk? Map every point where someone outside the organisation can stop a model, from the provider to the cloud to the governments that regulate them. Then qualify an open-weight model as a fallback, run it on infrastructure the organisation controls, keep a local record of which model ran, and rehearse switching over before it is needed. ### Does this mean critical systems should stop using hosted frontier models? No. Hosted frontier models remain the right tool for many jobs. The point is that a workload which must keep running needs a second path that does not depend on one provider, one cloud or one government's policy staying the same. The fallback should be tested on the real workload before it is ever needed. ## Sources - [When a Chatbot Becomes an Export](https://www.theregreview.org/2026/10/06/oktenli-when-a-chatbot-becomes-an-export/), The Regulatory Review - [Statement on the US government directive to suspend access to Fable 5 and Mythos 5](https://www.anthropic.com/news/fable-mythos-access), Anthropic - [Redeploying Claude Fable 5](https://www.anthropic.com/news/redeploying-fable-5), Anthropic - [The Department of Commerce Restricted Access to Anthropic's Latest Models. What Comes Next?](https://www.csis.org/analysis/department-commerce-restricted-access-anthropics-latest-models-what-comes-next), Center for Strategic and International Studies - [US saw risk of Anthropic models being diverted to foreign military intelligence](https://www.reuters.com/technology/anthropic-us-officials-meeting-monday-resolve-dispute-over-export-curbs-2026-06-15/), Reuters - [AI models have a 'kill switch' too: frontier access becomes a new front in export controls](https://futures.re.kr/en/articles/cm-349/), Futures - [UK retains early access to test latest US AI models](https://www.afp.com/en/uk-retains-early-access-test-latest-us-ai-models), AFP