On 12 June, access to frontier models inside ordinary workflows stopped. Anthropic removed Fable 5 and Mythos 5 after a government legal directive, then restored access in stages by 1 July. OpenAI's late-June preview of GPT-5.6 Sol also showed that government access processes can shape who uses a model and on what terms.
These events make a familiar dependency concrete. Model access can join the sanctions toolkit: not by freezing an account, but by withdrawing a capability teams have built into research, drafting, customer support and operations. The immediate issue is business continuity, not an abstract debate about AI.
An explicit AI policy should therefore include an access-cutoff plan. It should identify critical model dependencies, set ownership for continuity decisions and require data portability so prompts, evaluations, retrieval assets and workflows can move without starting again from zero.
Open-weight models are not a complete substitute for frontier services. They still require infrastructure, expertise and sound controls. But they are a credible fallback option, alongside multi-provider design and portable data. That is where digital sovereignty becomes operational: preserving the ability to keep working when access becomes leverage.
Sources