Estimated reading time: 12 minutes
This post examines proportionality and data centres from the attacker’s side. Once a commercial data centre is a military objective, the attacker must weigh the incidental harm that may be expected. Yet information on what the facility hosts is seldom easy to obtain (Klonowska and Schmitt, “The Challenges of Dual-Use Objects”).
My previous post found that a Party using a civilian data centre owes Article 58 of Additional Protocol I (AP I) protection only to civilians under its control. Other dependants, such as users in a neutral State, rely on the attacker’s obligations and on the law of neutrality (post #32, Section III).
The leading treatment is Akiyama’s. He asks whether disrupting identifiable services is reasonably expected to cause death, injury or damage. He accepts that some dependencies remain “uncertain or unknowable” before an attack (“Applying Reasonable Foreseeability”). He leaves open harm foreseeable in kind that cannot be traced to a named service.
My answer is that such harm counts. Death, injury or damage that a reasonable commander foresees in kind may be expected, even if no named user can be traced. Not knowing what a facility hosts does not, by itself, lower that expectation.
The assessment is made before the attack, on the information then available (post #5, Section 4). Foreseeable indirect effects reach systems that share infrastructure (post #17, Section III). Section I states the rules. Sections II to IV ask whether unnamed harm counts, what lowers the estimate, and what follows if none can be formed.
I. The rules: proportionality, precautions and what the attacker must know
Who and what the rules protect
The attacker’s duties come from AP I and its customary counterparts. An attack is an act of violence against the adversary (Article 49(1) AP I). The proportionality rule prohibits attacks that may be expected to cause excessive incidental harm. That harm is loss of civilian life, injury to civilians, damage to civilian objects, or a combination of these. It is weighed against the concrete and direct military advantage anticipated (Article 51(5)(b) AP I; Customary Law Study of the International Committee of the Red Cross (ICRC), Rule 14).
Article 57 adds the precautions, starting with constant care to spare civilians (Article 57(1)). Those who plan or decide upon an attack must do everything feasible to verify the target. They must take all feasible precautions in choosing means and methods to avoid or minimise incidental harm. They must refrain from excessive attacks, and cancel or suspend an attack once excess becomes apparent (Article 57(2)(a)(i)–(iii) and (b)). The customary rule spells out a duty to do everything feasible to assess whether an attack may be excessive (Customary Law Study, Rule 18). Violations by one Party do not release the other from these obligations (Article 51(8) AP I).
These rules protect every civilian. Article 50(1) AP I defines civilians by exclusion, and the 1987 Commentary stresses that nationality is not the important aspect (para. 1909). The attack rules apply in whatever territory an attack takes place (Article 49(2)). Nothing in Article 51(5)(b) or Article 57 limits the harm to civilians under a particular Party’s control. Users in a neutral State count like any other civilians.
The information standard and the harm that counts
The attacker is judged on the information available when it decides, and only death, injury and damage count. “Feasible” means practicable or practically possible in the circumstances ruling at the time (Germany, para. 2; United Kingdom, para. (b)). Germany declared that a decision is judged on “all information available to him at the relevant time”, not with hindsight (para. 4). The United Kingdom refers to “information from all sources which is reasonably available to them at the relevant time” (para. (c)). US doctrine states the same standard (US Department of Defense (DoD), Law of War Manual, § 5.3). Post #19 explains the reasonable-commander standard that applies it.
Expectation needs no certainty, but a mere possibility is not enough (Tallinn Manual 2.0, Rule 113, para. 12). Incidental harm includes foreseeable indirect effects, also called reverberating effects (Rule 113, paras 6–7; ICRC, 2023, at p. 2). Mere inconvenience does not count (DoD Manual, § 5.12.1.2). Whether loss of functionality is damage remains unsettled (Klonowska and Schmitt, “The Challenges of Dual-Use Objects”), and post #16 traced how far State positions diverge on data. I work within the narrower categories, as Akiyama does. An interrupted service counts only through the death, injury or damage it may be expected to cause.
II. Does harm count when no one can be named?
What “may be expected” asks
“May be expected” asks whether harm is expected. The words set no requirement that its victims be identified. The majority of the Tallinn experts read the test as binary. Once harm is expected, it enters the assessment without a discount for uncertainty. Otherwise it is “merely speculative” and does not count (Rule 113, para. 13). A minority would have scaled the required advantage to the probability of harm.
How likely must harm be? Robinson and Nohle reject the view that it must be more likely than not. Expected harm, on their reading, is harm likely to occur, and even a 40 per cent risk is foreseeable (at p. 118; doi:10.1017/S1816383116000552). The threshold remains contested, and my argument does not depend on fixing it. Foreseeability is also objective. Past practice, research, lessons learned and public information put commanders on notice (at pp. 121–122 and 145).
Akiyama applies this to data centres in two stages. The harm must be death, injury or damage. It must be reasonably expected on the information reasonably available at the time of decision (“Applying Existing Proportionality Principles”).
Harm foreseeable in kind
In my view, death, injury or damage that a reasonable commander foresees in kind may be expected, even where no named service can be traced. Harm foreseeable in kind is harm to the kinds of civilian users that the attacker knows, or can reasonably assume, the facility serves. Akiyama’s test reaches identifiable services and is silent on this case. My argument builds on his test and addresses that open question.
The ICRC supports the premise. Because civilian and military networks are highly interconnected, it expects incidental civilian harm from cyber attacks in most cases (ICRC, 2023, at p. 1). For data centres, it asks attackers to assess harm to populations that rely on them for essential services (ICRC, Frequently Asked Questions, 23 March 2026, question “What is the status and protection of data centres during armed conflict?”). Klonowska and Schmitt add that only foreseeable effects count, but that lost critical medical services would (“The Challenges of Dual-Use Objects”).
For kinetic attacks, Robinson and Nohle note that essential infrastructure is in most cases above ground and visible to a trained eye (at p. 133). A data centre’s dependants are not visible in that way. The cloud removes that premise, but the duty to weigh foreseeable harm remains.
The estimate in kind is therefore part of the expected harm. Only information the attacker actually has, for example about redundancy, can reduce it (Section III).
A comparison from the military-objective side
Targeting law already reasons by kind on the other side. An attacker may not know which road enemy forces will take. If use of a road in the network is reasonably likely, the network is a military objective (Tallinn Manual 2.0, Rule 101, para. 4). The harm side can be read the same way. Take State B from the scenario in post #32 as the attacker. It does not know which hospitals or utilities run on provider P’s facility in State N. It knows that the facility is a commercial data centre serving both in the region. Harm to patients and water users is then foreseeable in kind.
The comparison has limits. Discussing the purpose criterion, the International Law Association (ILA) Study Group requires information about a specific object, because each object must meet the definition individually (at p. 333). My thesis rests on the words “may be expected”, so it holds even for readers who follow the ILA.
The restrictive view
Henderson and Reece go further. They exclude indirect harm that remedial action may avoid, or that is “neither inevitable nor confidently predictable” (at p. 855). They concede that State practice and case law are too thin to settle the point (at p. 854).
As I see it, harm that follows from the loss of the service itself is not remote in the DoD sense. The Manual’s own example is a dependency case. Where destroying a power plant may be expected to kill patients in a connected hospital soon after, that harm counts (§ 5.12.1.3). A hospital that runs on a data centre is in the same position, even if the attacker cannot name it.
Henderson and Reece’s own example shows the limits of their view. Without water, they write, some civilians would likely die of thirst. They still treat these deaths as too remote, because aid, relocation or rain might intervene (at pp. 853–854). My conclusion is different. Where available information shows that remedial action will not reach part of the population, the resulting deaths may be expected and must be weighed.
III. What lowers the estimate
Information the attacker actually has can lower the estimate. Its ignorance cannot. Redundancy is the obvious case. Where data is stored across several data centres, striking one yields only a limited military advantage. Klonowska and Schmitt call this point particularly relevant to proportionality (“Targetable Objects Under IHL”).
Redundancy may also lower the expected harm. Akiyama counts redundancy, backup systems and alternative providers among the relevant information. A brief interruption of a service with effective redundancy may carry little risk of death, injury or damage (“Applying Reasonable Foreseeability”). Remedial action works the same way. Henderson and Reece expect no indirect harm where available information shows that civilians can bridge a water outage with bottled water (at p. 853).
Redundancy therefore cuts both ways. It lowers the military advantage and may lower the expected harm, but only where the attacker has information about it. Without such information, the estimate in kind is not reduced. If that estimate later proves too high, the decision is not unlawful for that reason. It is judged on the information available at the time, as the German and UK declarations confirm (Section I).
Nor may the attacker discount harm because the defender made it hard to see. A defender may run military and civilian workloads in the same facility, and a provider may keep its customers confidential. Neither entitles the attacker to discount harm it cannot trace to a named user but can reasonably expect. Violations by one Party do not release the other from its obligations (Article 51(8) AP I). Post #32 drew the same conclusion for a defender’s failure to take passive precautions (Section I).
IV. When no estimate can be formed
Outside the doubt rules
Sometimes the attacker cannot form even an estimate in kind. The doubt rules do not resolve that situation. Article 50(1) AP I addresses doubt about a person’s civilian status. Article 52(3) addresses doubt about whether an object normally dedicated to civilian purposes is used militarily. Post #20 found that residual doubt under either rule does not translate automatically into a duty to abstain (introduction; “Two scopes, not one”). Doubt about who depends on a facility is a third kind of doubt. Neither rule allocates it. The proportionality assessment and the precautions in attack govern it instead.
Civilian use and the duty to assess
Civilian use and civilian dependency are linked. Where the attacker knows, or can reasonably assume, that a data centre serves civilian users, harm to those who depend on it is foreseeable in kind. The ILA Study Group counts all foreseeable effects of an attack on the civilian use of an object that also serves military purposes as incidental harm (at p. 356). The ICRC, as noted, asks attackers to assess harm to populations that rely on data centres for essential services.
The attacker must then do everything feasible to assess whether the loss of that use may be expected to cause death, injury or damage (Customary Law Study, Rule 18). Commercial loss alone does not count, and the duty to assess does not by itself require abstention. Robinson and Nohle read this duty as including the proactive gathering of information where feasible, to an extent that depends on the operational context (at pp. 135 and 145). The 1987 Commentary states a duty to call for more information expressly for verifying the target (para. 2195). This post does not settle how far the duty to seek information about incidental harm reaches.
This does not ask the impossible. The duty reaches only what is feasible: what is practicable or practically possible in the circumstances ruling at the time, in the words of the German and UK declarations (Section I).
What the precautions require
Where no estimate in kind can be formed, Articles 57(2)(a)(ii) and (iii) AP I decide what the attacker may do. The precautions extend to indirect effects. The Tallinn experts require all feasible precautions to avoid or minimise indirect as well as direct harm, both in the choice of a cyber means and in how it is used (Tallinn Manual 2.0, Rule 116, para. 5).
For verification of the target, the experts go further. Where sufficient information cannot practicably be gathered, the decision-maker may have to refrain, modify the operation, or limit the attack to the components about which enough is known (Rule 115, para. 5). That passage concerns the target, not incidental harm. By comparison, it shows how a lack of information shapes what the precautions require.
The practitioner reading is that the attacker must choose means and methods whose effects it can estimate, or refrain where the harm that may be expected in kind would be excessive. Residual uncertainty does not by itself bar the attack. Where feasible precautions are not taken, however, incidental harm is more likely to be regarded as excessive (DoD Law of War Manual, § 5.12.3.2).
Conclusion
Existing law handles the opacity of cloud infrastructure without new treaty law. Opacity limits what the attacker can know. The duty to weigh the harm that may be expected stays the same. Three answers follow. Harm foreseeable in kind counts, even when no one can be named. Only information the attacker actually has can lower the estimate. Where no estimate can be formed, the precautions in attack decide, and they do not require automatic abstention.
This refines one point of post #17. There I proposed three analytical artefacts as the evidence base for the foreseeability judgement (“Second, insist on the foreseeability evidence base”). They remain good practice where feasible. The law, however, asks for an estimate in kind. It does not make mapping named dependants a precondition of a lawful attack.
Akiyama reaches a similar conclusion: attacks on data centres raise a question of applying existing law, not of making new law (Conclusion). For practitioners, that is workable. The law asks for good-faith judgement about kinds of harm, on the information available, within what is feasible.