International Initiatives and Core Principles for the Safe Use of (Military) AI: New Developments in 2025/2026

The most significant development in 2025–2026 was the move from general principles toward establishing implementation machinery. This includes capability-based thresholds, prerelease evaluations, testing and assurance regimes, procurement requirements, life-cycle risk management, human-control parameters, incident reporting, institutional governance, and crisis-risk reduction measures. The selected documents below reflect this shift and illustrate how the relevant arguments and mechanisms have evolved.

United States: A New Approach to Frontier AI Risks

Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence,¹ and the July 2025 document America’s AI Action Plan² marked a shift in overall U.S. policy toward the rapid development and adoption of AI. They direct agencies to identify and remove any federal measures regarded as barriers to AI innovation and to revise the previous administration’s implementation framework. This is not, however, tantamount to abandoning national security controls for the most capable models.

Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security,³ creates a more targeted security track. It directs the government to establish a classified benchmark process for advanced cyber capabilities and to use its results to designate “covered frontier models.” Developers may voluntarily provide the government with access to a covered model for up to 30 days before making it available to other trusted partners, thereby allowing prerelease testing while protecting confidential and proprietary information. The order expressly rejects a general mandatory licensing or government preclearance regime.

The order also proposes concrete cyber defense infrastructure around frontier models. It calls for an AI cybersecurity clearinghouse involving government, industry, and critical infrastructure actors to coordinate vulnerability discovery, validate findings, and prioritize remediation; expands access to AI-enabled defensive tools; and directs law enforcement agencies to prioritize cases of unlawful computer access or damage involving AI agents. The policy logic is therefore selective: broad regulatory restraint for AI generally, paired with capability-based intervention where frontier models materially affect national security cyber risk.

National Security Presidential Memorandum 11, Artificial Intelligence in the National Security Enterprise,⁴ is more operational than earlier statements of responsible use principles. It organizes national security AI policy around four pillars: adoption, adaptation, assurance, and accountability. Agencies are directed to use advanced commercial and open-source models where appropriate, adapt them to mission needs, and ensure that mission-dependent systems remain reliable, robust, steerable, and controllable. The implementation measures that follow from the memorandum are substantial. Among others, it orders an update of DoD Directive 3000.09, Autonomy in Weapon Systems,⁵ followed by recurring review; standardized national security testing, evaluation, verification and validation methods; a joint AI risk management and assurance strategy; procurement pathways able to onboard advanced models from multiple suppliers; secure computing infrastructure and a national security AI test range; and public-private partnerships for threat intelligence and AI red-teaming. It also creates mechanisms for incorporating nongovernmental AI expertise into national security work. The central policy proposition is thus accelerated fielding combined with a dedicated assurance, procurement, and security architecture.

European Union: Warning on Systemic Risks from Frontier AI

The European Systemic Risk Board’s June 2026 Warning on systemic cyber risks stemming from frontier artificial intelligence models⁶ makes a specific argument regarding the security consequences of capability gains. It considers frontier models capable of automating vulnerability discovery, exploit creation, and the execution of full cyberattacks at speed and scale. A central concern is the erosion of defensive time buffers: exploit development that previously took days or weeks may be reduced to minutes or hours, while a much larger volume of vulnerabilities could overwhelm patching processes and force institutions to choose between remaining exposed and deploying insufficiently tested patches. The EU document identifies four pressure points: time; defender capability; concentration/dependency on a small number of AI, cloud, and software providers; and the ability of authorities to calibrate testing and preparedness. It also highlights asymmetries between jurisdictions, attackers and defenders, and better- and less well-resourced institutions. It calls for adequate and proportionate European access to new frontier models, coordinated testing and cooperation, updated cybersecurity and stress-testing frameworks, and clear board-level governance and accountability.

Paris Declaration 2025: Maintaining Human Control in AI-Enabled Weapon Systems

The Paris Declaration on Maintaining Human Control in AI enabled Weapon Systems,⁷ issued at the February 2025 AI Action Summit and endorsed by 27 states, is a concise but unusually specific state-endorsed declaration on the boundary between AI-enabled weapons and human responsibility. It affirms that responsibility and accountability cannot be transferred to machines and commits endorsers to safeguards that preserve human judgement and control over the use of force. Its strongest formulation is that endorsers will not authorize “life and death” decisions to be made by an autonomous weapon system operating completely outside human control and a responsible chain of command. It further states that international humanitarian law (IHL) applies fully to AI-enabled weapon systems and that endorsers will not develop, deploy, or use systems that cannot be employed consistently with IHL. Compared with broader responsible-AI declarations, the document narrows the debate to an operational red line: exclusion of weapons operating entirely outside human control and command responsibility.

REAIM 2026: From Principles to Pathways to Action

The A Coruña Declaration – REAIM 2026 Pathways to Action⁸ deliberately shifts the Responsible Artificial Intelligence in the Military Domain (REAIM) process toward implementation. Its core proposition is “responsible by design”: legal, ethical, technical, and operational safeguards should be built into the military AI life cycle rather than added only at the point of deployment. The declaration specifically highlights risks arising from unanticipated or emergent behavior and the limits of AI explainability, and calls for robust testing, evaluation, verification, and validation (TEVV) to be integrated into relevant policies, doctrines, and procurement processes. The document also translates responsible use into organizational measures. It promotes training and risk management, continued monitoring and documentation, exchanges of practices, and—where compatible with national security—visits or exchanges involving relevant facilities and centers of excellence. At the international level, it suggests crisis communication arrangements to reduce misunderstanding, miscalculation, and unintended escalation; capacity-building through regional centers and knowledge hubs; and points-of-contact networks. For industry and academia, it emphasizes research on interpretability and robustness and calls on developers to address human-machine interaction and interface design proactively. Currently, the official REAIM page lists 45 endorsing states; neither the United States nor China are among them. The document remains politically nonbinding, but it specifies how earlier principles are to be translated into doctrine, procurement, testing, training, and international risk reduction practices.

United Nations: A Dedicated Military AI Governance Track and AI in Nuclear Command, Control, and Communications

United Nations General Assembly Resolution 80/58, Artificial intelligence in the military domain and its implications for international peace and security,⁹ is important because it structures a continuing UN process around military AI as a life-cycle governance problem. It covers predesign, design, development, evaluation, testing, procurement, deployment, use, sale, operation, and decommissioning; identifies risks including arms race dynamics, miscalculation, escalation, proliferation, bias, and use by non-state actors; and emphasizes safeguards involving human judgement and control over the use of force. The resolution also established a follow-up mechanism consisting of informal exchanges in Geneva on the Secretary-General’s report, existing and emerging normative proposals, and possible next steps, followed by a factual summary for the First Committee. It also encouraged Secretariat knowledge-sharing, as well as voluntary assistance and capacity-building, particularly for developing countries.

A separate new resolution is directly relevant to high-consequence military AI. General Assembly Resolution 80/23, Possible risks of the integration of artificial intelligence into command, control and communications systems of nuclear weapons,¹⁰ focuses on the possibility that AI could shorten decision windows, distort decision environments, or fail because of malfunction, intrusion, automation bias, or inaccurate outputs in nuclear command-and-control contexts. The resolution stipulates that human control and oversight be maintained over nuclear command, control, and communications systems, including those enabled by AI. It urges nuclear-weapon states to adopt and publish policies and doctrines that operationalize this requirement and explicitly provide that AI systems will not autonomously initiate decisions on the use of nuclear weapons. It also encourages voluntary information-sharing on implementation and the development of common understandings and confidence-building measures in relevant disarmament and treaty forums. This resolution is one of the clearest recent UN formulations of a specific human-control requirement for a defined military AI application.

CCW: Concrete Elements for an Instrument on Autonomous Weapons

The CCW Group of Governmental Experts has advanced materially beyond general exchanges of national positions. Its draft final report from September 2026¹¹ records a consensus set of elements for a future instrument, without prejudging whether that instrument will ultimately be legally binding. It characterizes a lethal autonomous weapon system as a combination of weapons and technological components capable of identifying, selecting, and engaging a target with force without intervention by the human operating the system, while excluding systems in which personnel determine the specific target or target group and the timing of attack. Substantively, the text sets out proposed prohibitions and control measures. It states that lethal autonomous weapon systems (LAWS) that are inherently indiscriminate, otherwise incapable of IHL-compliant use, or whose effects cannot be anticipated and limited as required by IHL must not be used, and reiterates the prohibition on directing attacks against civilians. Human judgement and control are operationalized through a responsible chain of command, human legal assessment, restrictions on target types, scale, duration, and geographic scope, human authorization for material changes to mission or target parameters, and the timely deactivation or neutralization of systems. It also calls for legal reviews, testing in representative environments, training, reliability/predictability/traceability, mitigation of data and automation bias, and internal mechanisms for incidents that may involve IHL violations.

Germany: Bundeswehr AI Concept

The German Federal Ministry of Defense (BMVg) KI-Konzeption: zukunftsfähige Verteidigung mit Künstlicher Intelligenz¹² is intended as a binding internal framework for the development, introduction, testing, and use of AI across the BMVg’s area of responsibility. Operationally, it presents AI as a means of enhancing information, command, and effect superiority; supporting multidomain operations and “mosaic warfare”; and improving staff and administrative processes. At the same time, it requires AI to be embedded within existing structures of organizational responsibility and data governance rather than treated as a standalone capability. The concept establishes a life cycle risk management and assurance regime. Testing with representative data is a prerequisite for authorization to use a system; high-risk systems are subject to “zero-failure” testing as a minimum requirement alongside quality metrics; material changes or retraining require renewed validation; and AI-specific minimum requirements are to be written into specifications, contracts, and acceptance procedures and subjected to independent review. Conformity review is required before first use and after material changes, while deployed systems are expected to support performance monitoring and, where relevant, revalidated in-mission updates. With regard to human control, personnel responsible for autonomous functions should understand system behavior and environmental interaction, evaluate reliability, set and validate rules of use, define mission objectives, target types, mission duration, and operational area, conduct Article 36 weapons reviews where applicable, and retain responsibility within a human chain of command. The concept establishes a coordinated governance structure around a Chief Artificial Intelligence Officer and CAIO Council, while retaining decentralized responsibility within existing processes.

NATO: Digital Architecture for Operational AI

The Alliance Digital Strategy,¹³ adopted by NATO in January 2026, goes beyond merely restating responsible use principles by locating AI within the Alliance’s future command-and-control and digital architecture. It calls for accelerated AI adoption in digital services; inference at the tactical edge to improve situational awareness, enable real-time risk-informed decisions, and increase operational autonomy; and AI-enabled sensor fusion, predictive analytics, and cognitive decision support for cross-domain command. These capabilities are linked to federated digital platforms, hybrid-cloud and tactical-edge infrastructure, standardized data and metadata, Zero Trust security, and software-defined defense. The strategy also sets out an implementation model: “adopt, buy, create” guides decisions on whether to reuse, procure, or develop digital capabilities; interoperability requirements and standards become particularly consequential when entities connect to federated networks for NATO-led operations; workforce development includes training on how to interact with AI-enabled systems; and progress is to be measured using the Digital Transformation Maturity Model and associated indicators. The new element is therefore not a separate AI code, but the integration of AI into the technical, organizational, and measurement architecture through which Alliance capabilities are fielded.

China: More Specific Principles and Risk Governance for Military AI

China’s Opportunities and Challenges Posed to International Peace and Security by the Application of AI in the Military Domain¹⁴ submitted to the UN in April 2025 provides a more detailed position on military AI than its earlier, more general governance statements. It argues that states—particularly major powers—should avoid seeking absolute military advantage through AI, prevent miscalculation and an AI arms race, and ensure that humans retain ultimate responsibility. It also states that weapon systems should remain under human control and that military AI must comply with international law, including IHL. The document further contains concrete governance proposals that are easily overlooked if it is treated merely as a statement of principles. China calls for forward-looking risk assessment; testing and evaluation based on the risk level of AI applications; tiered and categorized management; domestic legal, and regulatory systems and ethical guidelines; personnel education and training; and measures to reduce proliferation risks without obstructing peaceful innovation. Internationally, it emphasizes the UN as the principal forum and calls for policy coordination and capacity-building. The position therefore combines human-control language with a risk tiering and national implementation approach.

EU Artificial Intelligence Act: The Systemic Risk Regime Becomes Operational

For the EU Artificial Intelligence Act,¹⁵ the main new development is the implementation of its general-purpose AI (GPAI) regime. Since August 2025, providers placing GPAI models on the EU market have been subject to documentation, downstream information, copyright policy, and training-content transparency requirements. Providers of models already on the market before that date are subject to a later compliance deadline, and certain open-source models are exempt from some documentation requirements. Models with “systemic risk” face additional obligations: notification, standardized model evaluation, and adversarial testing; assessment and mitigation of Union-level systemic risks; serious-incident tracking and reporting; and cybersecurity protection for the model and its physical infrastructure. The regulation presumes high-impact capability above the statutory training-compute threshold, while also allowing Commission designation based on capability and impact.

The European Commission’s 2025 materials on General-purpose AI obligations under the AI Act,¹⁶ taken together, make these requirements more operational. The code’s “Safety and Security” chapter provides a voluntary compliance pathway for providers of systemic risk models, including safety and security frameworks, evaluation and risk management practices, model reporting, and post-market monitoring. Since August 2026, the Commission has been able enforce GPAI provider obligations,¹⁷ including through fines. This is relevant to a military AI annex not because the AI Act directly regulates systems used exclusively for military, defense, or national security purposes—it does not—but because it is becoming a detailed public regulatory model for capability identification, adversarial evaluation, systemic risk management, incident reporting, and security of frontier models.

Council of Europe: Ratification and Operationalization of Risk Assessment

The European Union ratified the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law¹⁸ in May 2026. The convention has not yet entered into force. It provides that parties shall establish iterative, graduated risk and impact assessments for actual and potential effects on human rights, democracy, and the rule of law; prevention and mitigation measures; transparency and accountability safeguards; procedural rights and remedies for affected persons; and the possibility of bans or moratoria on uses considered incompatible with those protected interests.

A 2026 follow-up element in the Council of Europe’s HUDERIA – risk and impact assessment of AI systems¹⁹ process is the approval of the HUDERIA Model: COBRA. The measures proposed by the HUDERIA framework are nonbinding, but provide states with a practical structure for identifying, assessing, and mitigating AI risks across the life cycle, using factors such as severity, probability, and reversibility, and for revisiting assessments as systems or contexts change. For military AI, its direct legal reach remains limited: the convention excludes matters of national defense and permits a national security carve-out subject to international law and democratic institutions. Its importance therefore lies in its use as a parallel model for operationalizing impact assessment, rather than as a direct military AI rulebook.