Frontier AI Is Becoming a State-Controlled Capability
The AI industry is entering a new phase. Some governments are no longer limiting themselves to regulating AI after deployment. They are beginning to influence whether frontier capability may exist, who may possess it, who may access it, and under what conditions it may operate. This represents a transition from AI governance to AI capability governance.
*This article primarily examines developments through a U.S. national security and governance lens, with selected European comparisons where they illuminate broader trends.
For several years, AI policy has focused on privacy, bias, transparency, safety, and accountability. Recent events in the United States suggest a significant shift in government posture. Rather than focusing solely on regulating AI after deployment, policymakers are increasingly influencing the availability of frontier capability itself. Recent events suggest the U.S. government is increasingly willing to intervene before frontier capabilities reach broad public deployment. Commercial solutions can be delayed, limited, or paused. The European approach differs. Rather than directly controlling frontier capability, the EU AI Act primarily regulates deployment, risk categories, and obligations placed upon AI systems. In addition, advanced cyber-capable models are increasingly being released only to trusted organizations. Capability is becoming licensed instead of universally available.
Recent actions involving frontier models demonstrate governments can effectively pause deployment while national-security questions are resolved. Export controls, custom AI chips hamstrung for export, cloud concentration, and compute restrictions are all examples of controls being enacted or discussed. Capability is increasingly controlled through infrastructure rather than software alone. Courts are now becoming part of frontier AI governance and the legal question is shifting from: “Can government regulate AI?” to “Can government determine who may possess frontier AI?”
The reason for this shift may not be immediately obvious. Frontier models increasingly resemble dual-use technologies. For an analogy, think of the encryption restrictions a few decades ago or for a more physical example, centrifuges used in enriching uranium for atomic weapons-devices that also have other, non-lethal uses. The same model that accelerates medical research or defensive cybersecurity can also accelerate vulnerability discovery, intelligence analysis, offensive cyber operations, or military planning. Governments have encountered this problem before with strong cryptography, satellite technology, and advanced manufacturing equipment, and now also begin treating frontier AI more like strategic infrastructure than consumer software.
State capability control is creating new architectural requirements. Government decisions do not directly control AI systems. They create policy requirements and those requirements must eventually be implemented through technical mechanisms and that transition, from public policy to enforceable system architecture, is precisely where AI Control Domains become useful. Policy cannot enforce itself, it ultimately depends on technical control points implemented within system architecture. Organizations need mechanisms for:
- identity
- authorization
- policy enforcement
- auditability
- provenance
- memory custody
- runtime monitoring
Control is moving from the model toward the surrounding architecture.
This is where AI Control Domains become useful. Threats describe what may happen and Control Domains describe where governance operates. As governments increase capability controls, architecture becomes the enforcement layer. Memory custody illustrates one architectural response to these emerging governance requirements. If governments require trustworthy operation, systems must demonstrate:
- controlled execution
- controlled memory
- verifiable deletion
- audit evidence
These become infrastructure capabilities rather than optional security features.
The defining question is no longer: “What can frontier AI do?” It is becoming: “Who decides who can use frontier AI, under what conditions, and how is that decision technically enforced?” This trend does not imply governments will ultimately succeed, it simply demonstrates that the locus of control is shifting from model developers alone toward governments, infrastructure providers, cloud operators, and the technical architectures that implement policy. Memory custody is but one example of how these architectural requirements may be implemented. EPAM/VEPAM represents one approach to implementing these requirements.
References
- OpenAI delay / government intervention https://www.reuters.com/legal/litigation/openai-defers-public-rollout-gpt56-us-seeks-early-access-frontier-ai-models-2026-06-26/
- EU AI Act https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- Anthropic Fable/Mythos suspension https://www.reuters.com/technology/us-blocks-foreign-access-anthropics-most-advanced-ai-models-axios-reports-2026-06-13/
Consulting: Need independent analysis or security support? See AI & Cybersecurity Consulting.