Anthropic’s decision earlier this year to restrict access to Claude Mythos, its new frontier model, reflects mounting trepidations over the system’s cybersecurity capabilities. Mythos is viewed as more than simply another model capable of producing polished prose or exceptional code. It can identify critical software vulnerabilities, reason through exploitability, combine weaknesses into viable attack chains, and expose flaws in widely used operating systems. Anthropic has understood Mythos as sufficiently advanced to warrant restricted access through trusted-access programs. Washington also imposed export controls restricting foreign-national access to Mythos on national-security grounds. Anthropic suspended access to the model entirely for all users until the Department of Commerce lifted the controls.
Models such as Mythos raise questions about access to frontier models and institutional capacity: who can act directly on what these systems discover, and who must rely on others to interpret, prioritize, and fix their vulnerabilities? Hence, Anthropic’s decision to withhold Mythos from the vast majority of users foregrounds how access restrictions, which seek to delimit risk, ironically raise second-order risks. This situation introduces the risk of concentrating defensive knowledge and capabilities, potentially amplifying cascading harms when such power is held by a narrow set of actors in the absence of robust AI governance frameworks.
This essay takes this conundrum as its analytical point of departure. As discussed below, this approach to Mythos offers an aperture to examine digital sovereignty in an era when control over advanced AI increasingly shapes who can discover, defend against, exploit, and govern critical digital infrastructure.
Digital Sovereignty and Interdependence
This advancement comes amid intensifying debates over digital sovereignty. For several years, digital sovereignty—a policy agenda predicated on a government’s capacity to develop, deploy, and govern digital technologies on terms it can meaningfully shape—has stressed data localization, compute, platform regulation, and institutional capacity-building. These discourses and ambitions unfurl within an asymmetric context in which much of the digital infrastructure that underpins Information and Communication Technology (ICT) sectors across the world, and particularly in the Global South, remains financially and technically dependent on U.S. and Chinese providers.
It is important to note that for many governments, the pursuit of digital sovereignty is inseparable from mounting perturbations over the weaponization of interdependence, which has revealed chokepoints that index both the successes of globalization and its vulnerabilities. By invoking digital sovereignty, governments seek to restructure domestic ICT sectors, reduce reliance on externally governed tools viewed as possible strategic liabilities, and, most saliently, ensure that the control of digital infrastructure allows emerging technologies to reflect local priorities, values, and development needs.
In this light, Mythos complicates calls for disentanglement, industrial policy, and national champions. For one thing, localizing data and infrastructure is believed to advance these pursuits of digital sovereignty. However, digitization itself hinges on interdependence: chips, cloud systems, vendor updates, and other externally supplied components and services. Accordingly, even domestically owned infrastructure may remain dependent on external systems of maintenance, diagnosis, and repair. Mythos makes this condition more visible. In an AI-mediated cybersecurity environment, digital sovereignty will rely not only on who owns data or infrastructure but also on how dependency is negotiated with allies and rivals. More fundamentally, it will increasingly turn on who controls frontier models—often developed or governed by leading U.S. and Chinese labs—capable of discovering, securing, exploiting, and helping remediate vulnerabilities embedded across expansive digital infrastructures.
Digital Chokepoints
Project Glasswing, Anthropic’s controlled-access cybersecurity initiative, brings these challenges of digital sovereignty and dependency into sharp view. It provides selected companies and institutions with early access to Mythos so they can identify and remediate vulnerabilities before comparable capabilities become more widely available. According to Anthropic, participating organizations can use Mythos Preview for defensive work, including vulnerability detection and patching automation.
Project Glasswing is, in many respects, a reasonable response to proliferation risk. If frontier models can accelerate vulnerability discovery, then software developers, cloud providers, financial institutions, and critical-infrastructure operators have a legitimate interest in gaining time to strengthen their immunity. A defensive head start may reduce the likelihood that newly discovered vulnerabilities are exploited before safeguards can be developed. Yet this move does not address the deeper governance questions: who is (and is not) granted access to these capabilities, and according to what technical, commercial, institutional, or geopolitical criteria?
Limited public transparency about the criteria for granting, prioritizing, or withholding access further reveals the challenges of dependency. Governments seeking to diminish reliance on externally governed digital systems may acquire greater cybersecurity capacity only through frontier models provided by firms they neither own nor govern. In this sense, advanced cybersecurity AI becomes a chokepoint: a strategically consequential capability through which access to defensive preparedness—and the technical and temporal advantage it can confer—can be selectively allocated.
Hence, Anthropic is doing more than selling a product. They are helping to determine when such capabilities are released, which actors receive access, under what conditions, and at what cost.
Leverage and Discretion
This chokepoint leverage raises critical matters. Cybersecurity has long been treated as a key modality of state sovereignty: governments defend networks, protect critical infrastructure, regulate software security, and, in some cases, stockpile vulnerabilities for intelligence or military purposes. Mythos complicates this arrangement. A leading AI laboratory may control a technology capable of discovering vulnerabilities more quickly than many state agencies, security researchers, cybersecurity firms, or criminal actors. Used responsibly and defensively, such a model could strengthen institutions and public safety. Used offensively, it could lower the barriers to sophisticated cyberattacks. Selectively withheld, even under reasonable concerns about proliferation, it could deepen asymmetries between those granted the capacity to protect themselves and those left comparatively exposed.
This challenge is especially acute in emerging markets, where banks, hospitals, power grids, mobile-money platforms, electoral databases, and government networks often operate through layered combinations of foreign hardware, proprietary software, cloud services, open-source dependencies, and outsourced maintenance. These arrangements already distribute technical authority unevenly: domestic institutions may use and regulate critical systems without possessing the expertise, source code, software inventories, dependency maps, institutional depth, or infrastructural control needed to secure them independently.
Kenya’s experience with its surveillance system illustrates this point. According to a government audit, local personnel had not fostered the technical expertise required to operate and maintain the Chinese-supplied system independently, leaving key functions dependent on the contractor; even the control-room interface was still in Chinese. Even more alarming, the audit identified weak access-control protocols that increased the risk of unauthorized entry, malicious code being introduced onto the server, and the misuse of the surveillance technology by unauthorized users.
The stakes of this chokepoint reveal that cybersecurity tools operate as modalities of power. If the most advanced defensive tools remain concentrated among a small number of frontier AI firms, allied governments, and preferred partners, cybersecurity capacity may become another instrument through which dependence is governed. States and institutions outside trusted circles may struggle to anticipate vulnerabilities and become increasingly reliant on external actors during crises. What appears as responsible risk management may also contribute to an asymmetry of preparedness, allowing powerful firms and government partners to influence the security posture of others by controlling capabilities they need but cannot easily reproduce.
The Challenge of Inclusion
Exclusion from trusted circles is not the only problem. Inclusion can also become a source of dependency. This dynamic is not entirely new. Earlier waves of digital transformation revealed a similar pattern: governments have often expanded their administrative and security capacities through U.S. systems they do not fully command—from Nvidia’s dominance in advanced AI chips to the hegemony of the U.S. dollar. These arrangements can produce real capacity. They allow governments to connect with citizens, monitor threats, digitalize public services, and participate in global digital markets. But they also embed state action within external architectures of finance, maintenance, standards, upgrades, and technical control.
In turn, Mythos extends this dynamic into the cybersecurity layer. It demonstrates that dependency can be produced not only through exclusion from advanced technologies but also through inclusion. Selected actors may gain access to powerful defensive capabilities, but the conditions of that provision may remain controlled by the model provider. Governments and firms become more capable, yet simultaneously more dependent on proprietary AI they neither own nor govern. Their security may improve, but the terms, duration, pricing, continuity, and distribution of that security are determined elsewhere. What they hold is not sovereign digital capacity but leased capacity.
This double bind suggests that countries need a coherent strategy to manage both the risks and opportunities of digital dependency. The aim cannot be to escape interdependence entirely. After all, even technological leaders such as China and the United States do not control the full digital stack. Rather, countries must negotiate from within their structural dependencies, delimit and mitigate their vulnerabilities, and use the opportunities afforded by building upon foreign platforms. That is what digital sovereignty should mean in practice: not the opiate of an untethered sovereignty, but the capacity to build leverage from within dependence.
Conclusion
The example of Mythos, then, clarifies the politics of digital sovereignty in an age of AI-enabled cybersecurity. The central problem is not simply whether governments have access to frontier models, nor whether provision should be unrestricted. Some limits may be necessary when systems can accelerate both vulnerability discovery and exploitation. The deeper question is how such limits are governed, who sets their terms, and how their benefits and burdens are distributed. When provision of cybersecurity AI is controlled by a small number of frontier labs and their state partners, defensive capacity itself becomes a chokepoint: a strategic capability that can be granted, delayed, priced, conditioned, or denied.
This matters because calls for digital sovereignty are, in part, responses to the weaponization of interdependence. Yet, the infrastructures through which AI operates cannot simply be disentangled from global circuits of chips, critical minerals, and cloud services. And so, the task is not to reduce digital sovereignty to a project of circumvention. Rather, it is to ask how interdependence is organized, who controls its chokepoints, and whether states and institutions can negotiate dependence in ways that strengthen mutual security and prosperity rather than reproduce vulnerability.
Concretely, this may mean that governments within trusted circles need to identify where critical dependencies actually reside and distinguish productive interdependence from vulnerabilities that structurally impede their capabilities. Procurement should then become a site of capability-building rather than outsourcing: governments can require greater transparency over software dependencies, local technical training, and viable exit or substitution pathways. Even so, these measures manage the form dependency takes rather than ending it. Frontier capability cannot simply be procured on better terms. Access ultimately remains something the provider can grant, condition, or withhold. However, for those outside trusted circles, the options are more limited. One avenue is to advocate through multilateral and multistakeholder forums—including the United Nations Global Mechanism on ICT Security and the Global Forum on Cyber Expertise—for mechanisms capable of governing the terms on which advanced cybersecurity capabilities are distributed, even though existing institutions do not directly govern the allocation of proprietary frontier capabilities by private firms. This limitation makes investment in domestic and regional technical capacity even more salient, including the pooling of capabilities—such as cybersecurity expertise and incident-response capacity—where national scale is simply insufficient.
In a word, the challenge is not exclusion, autarkic capacity-building, or concession to dependency, but rather the terms of interdependence. Mythos is not only a cybersecurity story. It is an analytical aperture into a wider geopolitical settlement in which the power to secure infrastructure—and the power to withhold that security—becomes central to the meaning of digital sovereignty itself.
