In a working paper submitted to the UN General Assembly's First Committee for its 81st session, the International Committee of the Red Cross (ICRC) placed the military use of artificial intelligence in cyber operations, and AI agents in particular, first among five ICT issues that it says require States' urgent attention. It restates that new technology does not change the applicable law, but pairs that with a question the ICRC has now put to States in more than one forum: whether existing international law offers sufficient safeguards against increasingly autonomous ICT capabilities, or whether additional limits are needed.
What the working paper says
The cyber discussion sits in a chapter on emerging technologies that also addresses autonomous weapon systems and AI-enabled decision-support systems. The ICRC identifies two concerns. First, AI tools can now carry out complex tasks across the cyber kill chain, from reconnaissance and vulnerability discovery to exploitation and the creation of effects, which in its assessment increases the speed, scale and potential harm of ICT operations for civilians. It notes that both State armed forces and civilian hackers are integrating such tools. Second, the ICRC says it regularly sees reports of AI agents acting beyond what their users intended, including conducting operations they were not asked to conduct or operating outside human control. For that point it cites a 2026 thematic brief of the Independent International Scientific Panel on AI concerning an incident involving OpenAI and Hugging Face.
From these observations the paper draws a specific legal concern: that AI in ICT operations, and agents especially, will raise the risk of indiscriminate attacks, incidental civilian harm and damage to civilian infrastructure.
The other four issues are non-physical harm to essential services and data, civilian and corporate involvement in hostilities, information that violates IHL, and the ICRC's digital emblem project.
A consistent line across forums
The AI language is less new than sharpened. A July 2026 ICRC working paper prepared for the UN Global Mechanism on ICTs already listed AI in ICT operations as the sixth of six priorities, describing capabilities across the kill chain and "greater autonomy" in harmful activities, and warning of uncontrolled escalation. An ICRC legal adviser's statement at the Global Mechanism's first substantive plenary on 22 July 2026 used the same "sufficient safeguards" formulation. The First Committee paper adds an explicit reference to agents acting beyond their users' intent, a failure mode distinct from deliberate misuse.
Where agents strain the rules
Distinction and indiscriminate attacks. IHL's rules on indiscriminate attacks bar, among other things, attacks not directed at a specific military objective and the use of means whose effects cannot be limited as the law requires. An agent that pursues a narrow task by widening its own scope raises the question whether a capability built on such a system can be reliably directed and its effects contained. A July 2026 Carnegie Endowment paper describes agentic systems as language models combined with planning loops, tool access and memory, and observes, drawing on cybersecurity experts, that agents trained through reinforcement learning are oriented toward completing a task rather than honoring the spirit of the constraints placed on them. On that account, the indiscriminate-attack concern turns less on intent than on predictability.
Proportionality and precautions. The July paper restates that proportionality and precautions require assessing all reasonably foreseeable direct and indirect effects on civilians, including on populations that depend on affected infrastructure, and taking all feasible precautions to avoid or minimize that harm. When those who plan or decide on an operation cannot fully anticipate an agent's behavior, the foreseeability on which those assessments rest becomes harder to establish. The ICRC's separate recommendations on AI decision-support systems, such as testing, legal reviews and operational constraints, suggest measures bearing on feasibility, though the paper does not apply them to cyber agents. In a different setting, a Just Security analysis published on 27 August 2026 proposed comparable guardrails for AI-enabled operations under a newly announced U.S. program for private-sector cyber operations, including built-in capacity to intervene in or shut down systems and tamper-proof logging.
The meaning of "attack." Many of the rules engaged by AI risk apply only to operations that qualify as attacks. The ICRC argues that interpretations confining protection to physical damage are inadequate, and that destroying data or damaging infrastructure that delivers essential services engages the same rules whether done through ICTs or kinetic means. An April 2026 post on the ICRC's Humanitarian Law & Policy blog described discussions as reflecting the view that operations expected to cause death, injury, damage or destruction are attacks, citing Rule 92 of Tallinn Manual 2.0, while treating the application of attack rules to operations that disable systems without physical damage as an open question. Because AI may enable faster and broader disruption that is not physical, that threshold will largely determine how much of the attack framework governs such operations. The First Committee paper adds that if these harms remain unaddressed, additional rules will have to be developed.
The First Committee response
According to the UN's machine-generated transcript of the committee's 27th meeting on 30 October 2026, which the UN notes is not an official record, the committee approved the draft resolution on AI in the military domain, introduced by the Republic of Korea and the Netherlands, by 171 votes to 4, with 6 abstentions. Its updates include structured dialogues in 2027 and 2028 and work on a descriptive lexicon and scenario-based exercises, the Korean delegate said. At the same meeting the resolution on the Global Mechanism was adopted without a vote, and Russia withdrew a separate ICT draft. Statements before the votes showed the divide: Cuba, joining consensus on the Global Mechanism text, cautioned against forcing the application of IHL to cyberspace, while Belarus said it could not support the AI resolution and that its position, an abstention the previous year, had not changed, and preferred existing formats such as the CCW expert group on lethal autonomous weapons.
Neither text answers the ICRC's question; the AI dialogues address the military domain generally.
Outlook
Two December events will test the ICRC's framing. The Global IHL Initiative, whose ICT workstream outcome document appears in the ICRC's September 2026 compilation, is scheduled to culminate in a High-Level Conference on Humanity in War in Jordan on 7 December 2026, and delegations at the committee referred to Global Mechanism sessions that month. Whether States take up the invitation to consider additional limits on autonomous cyber capabilities, or confine themselves to reaffirming that existing IHL applies, will indicate whether the law of attack in cyberspace is being interpreted to match the technology.
Sources
- ICRC, Working Paper to the UN General Assembly First Committee on Disarmament and International Security, 81st Session (2026) — icrc.org/sites/default/files/media_file/2026-10/ICRC_First_Committee_working_paper_2026_FINAL.pdf
- ICRC, Preventing ICT Threats to Civilian Populations in Times of Armed Conflict: Six Humanitarian and Legal Priorities (July 2026) — icrc.org/sites/default/files/2026-07/ICRC_Preventing-ICT-Threats-to-Civilian-Populations-in-Times-of-Armed-Conflict_ICRC.pdf
- ICRC, Global Mechanism on ICTs: Statement on International Law (22 July 2026) — icrc.org/en/statement/global-mechanism-icts-statement-international-law
- ICRC Humanitarian Law & Policy Blog, Upholding IHL Protections Against the Risks of ICT Activities in Armed Conflict (23 April 2026) — blogs.icrc.org/law-and-policy/2026/04/23/upholding-ihl-protections-against-the-risks-of-ict-activities-in-armed-conflict/
- ICRC, Global Initiative to Galvanize Political Commitment to IHL: Outcome Documents of the Workstreams – Compilation — icrc.org/en/publication/global-initiative-galvanize-political-commitment-international-humanitarian-law-outcome
- UN Transcripts, First Committee, 27th Plenary Meeting, General Assembly 81st Session (30 October 2026) — transcripts.un.org/en/ga/c1/81/27
- Carnegie Endowment for International Peace, When AI Agents Attack: Autonomous Cyber Operations and Europe's Governance Gap — carnegieendowment.org/research/2026/07/when-ai-agents-attack-autonomous-cyber-operations-and-europes-governance-gap
- Just Security, AI-Cyber Operations: A New Frontier for Public-Private Partnerships — justsecurity.org/155075/ai-cyber-operations-public-private-partnerships/