On 25 September 2026, the US Federal Court of Appeals rejected the artificial intelligence (AI) company Anthropic’s challenge to its designation by the Pentagon as a “supply chain risk” on national security grounds. The decision, adopted by a two-to-one vote, marks an important stage in the months-long dispute between Anthropic and the US Department of Defense, which stems from the safety restrictions concerning the military use of the company’s Claudemodels.[i] It would be inaccurate to regard this tension merely as a contractual dispute between a public institution and a technology provider. The integration of AI into different areas of national security, from intelligence analysis to operational planning, raises the question of how the relationship of authority and control between the state and technology providers will be shaped in an environment where a significant part of strategic technological capacity is developed by private companies.
Behind the tension between the Pentagon and Anthropic lies the extensive cooperation previously established between the two actors. In 2025, the Pentagon’s Chief Digital and Artificial Intelligence Office (CDAO) awarded Anthropic an agreement worth up to $200 million to develop advanced AI capabilities for critical national security challenges in the warfighting and enterprise defence domains.[ii] According to Anthropic, Claude models are used in a range of national security tasks, including intelligence analysis, modelling and simulation, operational planning and cyber operations.[iii]The current dispute therefore does not stem from the company’s general opposition to the military use of AI, but from a narrower yet strategic debate over who will determine the limits of its areas of use.
The tension became more pronounced when the Pentagon demanded that AI providers open their models to all lawful national security uses. Anthropic, however, refused to lift its restrictions in two areas of use: mass domestic surveillance and fully autonomous lethal weapons. The company argues that current advanced AI models are not sufficiently reliable to select targets and use lethal force without human intervention, while contending that AI-enabled mass surveillance poses significant risks to fundamental rights.[iv] For the Pentagon, on the other hand, the possibility that the functions of a strategic technology used in national security missions could be restricted by a private provider creates a different kind of risk in terms of military decision-making and operational control.
For this reason, viewing the Pentagon-Anthropic dispute as a “conflict between security and ethics” would be incomplete. The parties’ positions point rather to a divergence over how two different security risks should be prioritised. While Anthropic focuses on the risks that may arise from the use of models whose reliability has not been sufficiently proven in certain military and surveillance activities, the Pentagon regards the dependence of national security capacity on the technical and contractual preferences of a private technology provider as the problem. The majority opinion in the court ruling of 25 September, too, rather than rejecting outright the AI safety risks put forward by Anthropic, left room for the decision on how these risks are to be balanced to be assessed within the framework of the state’s national security powers. The debate thus moves beyond the question of whether AI is safe and extends to the distribution of the authority to determine the limits of use of strategic technologies.
The significance of the Pentagon-Anthropic dispute becomes more evident as the US Department of Defense increasingly develops its AI capacity through partnerships with private technology companies. According to research published by the Center for Strategic and International Studies (CSIS) in September 2026, approximately 83 per cent of the $4.1 billion in obligations undertaken by US federal agencies under AI-related contracts between fiscal years 2019 and 2025 belongs to the Department of Defense.[v] This picture shows that the Pentagon has become a major buyer in the field of AI and that its relationship with private technology providers is becoming increasingly important for national security. Consequently, the dispute with Anthropic goes beyond a contractual debate with a single company and raises the question of how the state will manage its dependence on strategic technologies developed by the private sector.
This transformation also broadens the scope of the concept of defence supply chain security. Traditionally, supply chain risks have been assessed in terms of access to critical physical components, dependence on foreign manufacturers, sabotage or the reliability of technological hardware, whereas AI systems create a different form of dependence. Since a model’s functions, safeguards and usage parameters can be continuously updated by the software provider, the reliability of the supply chain is not limited to whether the technology reaches the state. The conditions under which the capacity accessed by the state can be used also become part of the relationship established with the provider. In this respect, the Pentagon’s assessment of Anthropic as a supply chain risk shows that defence procurement security can expand from the continuity of physical components to the software provider’s capacity to control the strategic system.
These characteristics of AI systems partly distinguish advanced AI companies from traditional defence contractors. Traditional defence companies have also long played an important role in developing states’ military capabilities. However, the continuously updated, software-based nature of advanced AI models can give provider companies an ongoing sphere of influence over the technology’s functions and safety parameters. States retain their fundamental decision-making capacity through contractual terms, procurement policies, regulatory instruments and national security powers. Nevertheless, the concentration of the development of critical technological capacity in the private sector increases the strategic importance of the relationship that the state establishes with private actors in order to be able to use this capacity.
The Pentagon’s simultaneous work with different advanced AI providers can be regarded as a means of managing this dependence. In addition to Anthropic, the Department of Defense has also signed agreements with Google Public Sector and other advanced AI providers for the development of national security applications.[vi] Including multiple providers in the defence ecosystem has the potential to reduce the risk that a single company’s technological or contractual preferences could turn into a critical bottleneck for the state. Competition in AI procurement policy may thus be shaped not only by which model offers greater capability, but also by the extent to which the state can preserve its operational autonomy and technological alternatives.
That said, interpreting the Pentagon-Anthropic tension as a sign that private technology companies have turned into independent actors capable of determining national security policies would also be an overreach. While Anthropic has the capacity to set its own terms of use, the Pentagon is able to change the company’s position in defence contracts and turn to alternative providers. The emerging structure therefore points not to the disappearance of state authority, but rather to a new form of interdependence arising from the concentration of the production of strategic technological capacity and of the decision-making authority over its political and military use in different actors.
In conclusion, the Pentagon-Anthropic tension shows that the integration of AI into national security systems is not merely a matter of new military capabilities. The significant role that private companies play in the development of strategic AI technologies calls for a reassessment of the concepts of supply security, operational control and technological dependence. While the court ruling of 25 September 2026 has made the legal dimension of this transformation visible, the underlying structural issue is how the relationship between the owner of technological capacity, its provider and the actor that determines its limits of use will be established in the age of AI. In this context, the key question for future national security policies will be not so much whether states can access the most advanced AI systems, but how they can simultaneously preserve operational autonomy and the reliable use of technology while managing their dependence on these systems.
[i] “US Appeals Court Upholds Pentagon’s Blacklisting of Anthropic”, Reuters, https://www.reuters.com/world/us-appeals-court-declines-block-pentagons-blacklisting-anthropic-2026-09-25/, (Accessed: 27.09.2026).
[ii] “Contracts For Jul. 14, 2025”, U.S. Department of War, https://www.war.gov/News/Contracts/Contract/Article/4243673/contracts-for-jul-14-2025/, (Accessed: 27.09.2026).
[iii] “Statement on our discussions with the Department of War”, Anthropic, https://www.anthropic.com/news/statement-department-of-war, (Accessed: 27.09.2026).
[iv] Ibid.
[v] Yasir Atalan, Erik Tiersten-Nyman and Benjamin Jensen, “Understanding the U.S. Federal Government’s AI Spending”, Center for Strategic and International Studies (CSIS), https://www.csis.org/analysis/understanding-us-federal-governments-ai-spending, (Accessed: 27.09.2026).
[vi] U.S. Department of War, op. cit.
