In November 2025, Anthropic disclosed a cyber espionage campaign in which its Claude Code system reportedly targeted about 30 targets worldwide. According to the company’s investigation, AI performed an estimated 80-90 per cent of the campaign’s operations, while human operators intervened at only a few critical decision points.
Anthropic assessed the activity as originating from a Chinese state-sponsored group. However, this remains the company’s intelligence assessment, not a legally established attribution of conduct to China. The incident nevertheless illustrates a significant shift: AI systems can increasingly execute substantial portions of sophisticated cyber operations with limited human intervention.
The implications extend beyond cybersecurity. When an AI system identifies vulnerabilities, develops exploits, extracts information, and determines successive operational steps, a fundamental legal question emerges: Does increasing machine autonomy change how international responsibility is established?
The answer is not that international law has become obsolete. Instead, AI highlights the difficulties of applying established rules of State responsibility to operations where human decision-making is less visible, technical evidence is distributed across private infrastructures, and the boundaries between State and non-State conduct are increasingly hard to draw.
Autonomy Does Not Eliminate Attribution
Under international law, State responsibility arises when conduct attributable to a State breaches an international obligation.
The International Law Commission’s Articles on Responsibility of States for Internationally Wrongful Acts (ARSIWA) provide the principal framework. Article 4 concerns conduct attributable to State organs, while Article 8 addresses persons or groups acting on a State’s instructions or under its direction or control.
AI systems are not independent legal actors within this framework. They are technological instruments through which human actors and institutions may operate.
Consider an intelligence agency deploying an autonomous AI agent to identify vulnerabilities, access foreign networks, and extract information. The system may select technical methods or execute intermediate steps without continuous human approval. Yet delegating these functions to an algorithm does not, by itself, remove the agency’s relevant conduct from the scope of Article 4.
This does not mean that every consequence the system generates automatically creates international responsibility. Attribution and wrongfulness remain distinct legal questions.
A cyber operation attributable to a State is not necessarily an internationally wrongful act. Cyber espionage, for instance, is not automatically prohibited as a distinct category under general international law, although particular methods or effects may violate applicable international obligations.
Similarly, foreseeability does not, by itself, establish responsibility under ARSIWA. The significance of foreseeable consequences depends on the content of the particular primary obligation allegedly breached.
Operational autonomy therefore does not confer legal independence. AI may change how State activities are executed, but it does not automatically alter the legal framework through which those activities are assessed.
Non-State Actors and the Limits of Control
The more difficult question arises when AI-enabled cyber operations are conducted by non-State actors.
Under Article 8 of ARSIWA, identifying links between a group and a State is not necessarily sufficient to attribute the group’s conduct to that State. The relevant instructions, direction or control must be established in relation to the conduct at issue.
The International Court of Justice addressed this distinction in its landmark Nicaragua v. United States judgment (1986), emphasizing the importance of effective control over the operations in which the alleged violations occurred. The Court later reaffirmed this standard in its Bosnian Genocide judgment (2007).
AI-enabled operations create practical difficulties in applying these established principles.
In conventional cyber campaigns, communications between operators, instructions, financing arrangements and operational coordination may provide evidence of the relationship between a State and a non-State group.
An agentic AI system, however, may perform multiple technical tasks following limited initial instructions. Activities previously requiring continuous human direction can be delegated to automated workflows.
This does not necessarily mean that AI produces less evidence. Model interactions, API requests, authentication records and system logs may generate extensive technical traces.
The challenge is whether these traces can establish the legally relevant relationship between a State and the conduct in question.
Identifying the infrastructure used in an operation is not equivalent to identifying its operator. Identifying an operator is not equivalent to establishing State control. And establishing political or operational connections does not automatically satisfy the requirements of legal attribution.
AI does not remove these distinctions. It makes their careful application increasingly important.
Due Diligence: A Separate Question of State Responsibility
When the conduct of a non-State actor cannot be attributed to a State, a different question may arise: whether that State has breached an obligation of its own by failing to take appropriate measures.
This is where the concept of due diligence becomes relevant.
The 2015 United Nations Group of Governmental Experts report (A/70/174), in paragraph 13(c), endorsed a voluntary norm that States should not knowingly allow their territory to be used for internationally wrongful acts involving information and communications technologies.
However, this non-binding norm must be distinguished from an independently binding obligation under customary international law.
States continue to disagree over the existence and scope of a general cyber-specific due diligence obligation.
The disagreement is visible in the United Nations’ 2021 Official Compendium of National Contributions (A/76/136). For example, Norway holds that due diligence obligations apply to harmful cyber operations and may require reasonable measures when a State knows or should know about relevant activities. The United Kingdom, by contrast, has maintained that State practice is insufficient to establish a specific customary due diligence rule applicable to cyberspace.
AI adds complexity to this already contested field.
An operation may involve a model developed in one jurisdiction, cloud infrastructure situated in another and an operator acting from a third State. The mere presence of a model provider or data centre within a State’s territory does not automatically establish that State’s international responsibility.
Where an applicable due diligence obligation exists, the relevant assessment concerns the connection between the harmful activity and the State, the knowledge available to its authorities, the measures reasonably within their capacity and the requirements of the particular obligation.
These questions become harder when the same AI capability can be used for legitimate security testing, research, and malicious operations.
The central challenge, therefore, is not simply whether States should prevent harmful uses of AI. It is how knowledge, jurisdictional connections, technical capacity and reasonable preventive measures can be assessed without turning due diligence into an unlimited obligation to monitor or control technological activity.
Importantly, due diligence cannot serve as a shortcut around attribution requirements. Where legally applicable, it concerns a potentially separate breach arising from the State’s own action or omission, rather than automatic responsibility for the underlying conduct of another actor.
The New Geography of Evidence
The Anthropic incident reveals another significant development: private technology companies may possess some of the most relevant evidence concerning AI-enabled cyber operations.
In its November 2025 disclosure, Anthropic detected suspicious activity, investigated the campaign, disabled associated accounts, notified affected organisations, and coordinated with authorities.
The company also reported that AI executed much of the campaign’s operational work, while humans remained necessary at certain critical stages.
Such findings illustrate the growing importance of model providers, cloud operators and other private technology companies in reconstructing cyber incidents.
These actors may possess information about prompts, authentication records, system interactions, operational sequences and the points at which human operators intervened.
However, technical attribution and legal attribution are not interchangeable.
A private company’s assessment of an operation’s origin, sophistication or suspected sponsor does not establish State responsibility under international law. Technical indicators may support a broader evidentiary assessment, but they do not replace the legal requirements governing attribution and wrongfulness.
International law also does not set a universal evidentiary threshold for every attribution decision. The requirements of judicial proceedings must be distinguished from States’ technical assessments and decisions to make public attribution statements.
The emerging challenge is therefore not merely the quantity of evidence available. It is how relevant evidence can be preserved, authenticated, shared and assessed while respecting privacy, commercial confidentiality, national security and other legitimate legal interests.
AI-enabled operations create a growing asymmetry: States remain central to the law of international responsibility, while private actors may control information essential to its application.
The solution cannot be to transfer legal attribution decisions to technology companies. Instead, it requires stronger cooperation mechanisms and reliable evidence-sharing without confusing private threat intelligence with legal determinations.
Applying Existing Rules Before Creating New Ones
The increasing autonomy of AI-enabled cyber operations does not necessarily require an entirely new framework of international responsibility.
More immediate progress could be achieved in three areas.
First, States should articulate more clearly how existing rules of attribution and relevant international obligations apply to AI-enabled cyber activities.
The United Nations’ Global Mechanism on ICT Security, which held its first substantive plenary in July 2026, provides an institutional forum for continuing discussions on international law, responsible State behaviour and cooperation in cyberspace. Although it is not specifically an AI governance mechanism, these discussions may inform future consideration of AI-enabled operations.
Second, technical governance should support legal accountability. Appropriate logging, traceability, incident reporting and human oversight can improve the ability to investigate operations involving autonomous systems.
These safeguards should not be confused with existing, universally applicable international legal obligations. Their value lies in supporting transparency, risk management, and evidence preservation.
In a September 2026 UNIDIR study, Beyza Ünal examines how lessons from voluntary norms of responsible State behaviour in cyberspace could inform governance discussions concerning AI in the military domain. The study identifies cooperation, transparency, incident reporting and private-sector engagement as promising areas for further development, while emphasizing that cybersecurity norms cannot simply be transferred to military AI without adaptation.
Third, States and technology companies should develop clearer procedures for preserving and sharing evidence relevant to AI-enabled cyber incidents. Such arrangements should balance investigative needs with data protection, confidentiality, security and procedural safeguards.
These measures would not eliminate disagreements over cyber due diligence or resolve every attribution challenge. They could, however, improve the practical conditions under which existing rules are applied.
Responsibility Without Continuous Human Control
AI is transforming the operational structure of cyber activities. Increasingly autonomous systems can accelerate reconnaissance, vulnerability exploitation and data processing while reducing the need for continuous human intervention.
However, reduced human intervention should not be mistaken for the disappearance of legally relevant human or institutional conduct.
The distinction between attribution and wrongfulness remains fundamental. So does the distinction between responsibility for the conduct of State organs, responsibility for operations carried out by non-State actors under State instructions or control, and responsibility arising from a State’s own failure to comply with an applicable obligation.
The problem is not that international law has no place for AI-enabled operations. Increasingly autonomous activities challenge the availability, interpretation, and assessment of the evidence through which existing legal rules operate.
AI does not create a vacuum of legal responsibility. It exposes the practical and evidentiary limits of a framework that still distinguishes between conduct, control, knowledge, and internationally wrongful acts.
The priority is therefore not to attribute international responsibility to algorithms. Instead, it is to clarify how States’ conduct and obligations can be assessed as operational decisions become increasingly automated and the information needed to evaluate them is dispersed across public and private actors.
References
Anthropic. (2025, November 13). Disrupting the first reported AI-orchestrated cyber espionage campaign. Official publication.
International Law Commission. (2001). Articles on Responsibility of States for Internationally Wrongful Acts, with Commentaries. Particularly Articles 2, 4 and 8. Official UN text.
Ayşe Özge Erceiş is a PhD candidate in Public International Law at Université Paris-Panthéon-Assas, where her research focuses on cross-border transfers of sensitive State data, cybersecurity, and State responsibility. She holds an LL.M. in International Law and International Organizations from Université Paris 1 Panthéon-Sorbonne and is a member of the Istanbul Bar Association. She is the founder and editor of Cyber Juris Weekly, a publication dedicated to international cyber law and digital governance. Her work explores the intersection of international law, emerging technologies, and international security.
