Estimated reading time: 11 minutes
Armed forces increasingly run data and services on commercial cloud infrastructure that hospitals, utilities and ordinary users also rely on. When a force uses a civilian data centre for military purposes, that facility can become a military objective. The precautions against the effects of attacks in Article 58 AP I then come into play. They require the using Party to protect civilians under its control against the resulting dangers. This post asks a narrow question: whom does that duty protect? My answer is that Article 58 protects only the civilians and civilian objects under the using Party’s control. Control is tested for each person and object, wherever the servers sit.
Earlier posts prepare the ground. Post #19 treated Article 58 as a duty in its own right, distinct from the attacker’s duties under Article 57. The post on precautions in cyber attacks under Article 57 AP I took the attacker’s side. Post #31 asked which neutrality rules apply to AI compute in neutral territory. This post stays with international armed conflict and with cloud services and data centres. It states the legal test. It does not assess the lawfulness of any ongoing operation, whichever party is involved.
I. The rules: Article 58 AP I and its customary counterparts
Article 58 belongs to Additional Protocol I (AP I), which applies in international armed conflicts (Article 1(3)). AP I has 174 States parties, according to the International Committee of the Red Cross (ICRC). Other States are bound by the same duty as customary law. The ICRC Customary IHL Study restates it as Rule 22, for international and non-international armed conflicts alike. For cyber operations, most Tallinn Manual 2.0 experts limited the rule to international armed conflict (Rule 121, para. 2). Austria applies Article 58(a) and (b) subject to the requirements of national defence. Its reservation does not cover sub-paragraph (c) (ICRC treaty database). Switzerland withdrew its reservations to Articles 57 and 58 in 2005 (AS 2005 2619).
Article 58 obliges the Parties, “to the maximum extent feasible”, to do three things. They must endeavour to remove civilians and civilian objects under their control from the vicinity of military objectives (a). They must avoid locating military objectives in or near densely populated areas (b). And they must take “the other necessary precautions” to protect civilians and civilian objects under their control (c). These precautions against the effects of attacks are often called “passive precautions”. The qualifier governs all three limbs. Many States read “feasible” as “practicable or practically possible”, weighing humanitarian and military considerations (Rule 22). As the 1987 ICRC Commentary puts it, “no one can be required to do the impossible” (para. 2245, p. 692).
The duty covers only civilians and civilian objects “under their control”. The Commentary ties Article 58 to measures a Power takes at home “or in territory under its control” (para. 2239, p. 692). Yet control is not purely territorial. Eric Talbot Jensen reports that the drafters chose “control” over “authority” to reflect the de facto situation (98 IRRC 147 (2016), at p. 162; DOI 10.1017/S1816383117000017). The Tallinn experts divided over its meaning for cyber infrastructure, a split Section III takes up. The German and US manuals speak of responsibility rather than control (ZDv A-2141/1, para. 412; DoD Law of War Manual, § 5.14).
Three neighbouring rules keep Article 58 in its lane. Shielding military objectives with civilians is prohibited, and the prohibition turns on that purpose (Article 51(7)). A Party’s failure to take passive precautions does not release the attacker from its own obligations (Article 51(8); Rule 121, para. 14). Special rules protect medical units, objects indispensable to survival and works containing dangerous forces (Articles 12(4), 54(2) and 56(5)).
An object is a military objective if its nature, location, purpose or use makes an effective contribution to military action. Its destruction or neutralisation must also offer a definite military advantage (Article 52(2) AP I). There is no third, “dual-use” category (Tallinn Manual 2.0, Rule 101, para. 1). An attack is an act of violence against the adversary (Article 49(1)). A cyber attack is a cyber operation reasonably expected to cause injury, death, damage or destruction (Rule 92). The cyber rule on passive precautions, Rule 121, rests on Article 58(c) as a catch-all (paras 1 and 4). Article 58 matters here for one reason. Military use can turn a shared data centre into a military objective and expose the civilians who depend on it.
II. Using and sharing civilian cloud infrastructure: what is settled and what is not
The first question is whether a Party to an armed conflict may use civilian cloud infrastructure for military purposes at all. The answer is clear: it may. The use has a price, because the facility may become a military objective. But the use itself is lawful. The Tallinn experts treat military use of civilian objects as distinct from targeting, much like taking over public roads (Rule 101, para. 9). On wholesale use, even the most restrictive voice agrees, so the treatments converge. David Koplow accepts that overt, wholesale use turns a civilian object into a military objective. “This sort of military occupation of a formerly civilian object is not illegal” (13 Harv. Nat’l Sec. J. 25 (2022), at p. 105). The Tallinn experts go further. They found that State practice establishes the legality of using cyber infrastructure for both military and civilian purposes (Rule 121, para. 13). That is where agreement ends.
Views part on the second question: may a force share one facility with civilian users? The restrictive view, at its strongest, runs as follows. Koplow writes about satellites, but his reasoning applies to shared data centres by analogy. He treats partial or time-limited military use of a civilian satellite as a breach of Article 58. The same goes for renting commercial ground stations (at pp. 106–107). Yet he accepts that genuine military necessity can justify intermingling (at p. 114). The ICRC takes a policy line in the same direction. It warns that growing military use of civilian infrastructure, cloud included, raises the risk that civilians are targeted. This trend, it says, “must be reversed” (statement to the UN Open-Ended Working Group, 13 December 2023). The ICRC’s Global Advisory Board recommended segmenting military from civilian data and communications infrastructure where feasible (final report, Recommendation 11, p. 13). A State using a non-segmented commercial cloud “could be regarded as deliberately shielding” its assets (ibid.). The report reflects a consensus of the members, not necessarily of their organisations (p. 6).
The opposing view rests on military necessity and State practice. John Goehring argues that feasibility must weigh military necessity, which in his view justifies intermingling in space (Lieber Institute, “Assessing Feasibility”; “The Case for Lawful Intermingling”). Without evidence that States oppose intermingling on legal grounds, he finds claims of illegality premature (“The Importance of State Practice”). He calls Koplow’s approach “a single factor test in which intermingling equates to illegality” (“Taking Necessity Seriously”). Koplow’s own necessity exception goes unmentioned. Svenja Berrang, writing in a personal capacity, takes a cautious line. She suggests that “one could argue that it is neither practical nor practicably possible” to separate military and civilian space systems completely (ICRC Law & Policy blog). Klonowska and Schmitt point to separation measures that may become increasingly important for data centres (Just Security). The ICRC asks States and the tech sector to work towards a common understanding of limits on the military use of civilian digital infrastructure (same statement).
One point from the restrictive side needs an answer. As I see it, using a non-segmented commercial cloud is not, by itself, shielding. Article 51(7) targets the use of civilians “in attempts to shield military objectives from attacks”. It turns on that purpose. Co-location without it may fall short of Article 58, but it is not shielding.
I leave the sharing dispute open here. The question this post addresses holds whichever view prevails. When a Party does rely on a shared facility, whose civilians does Article 58 protect?
III. Whom does Article 58 protect when the cloud crosses borders?
Control attaches to people and objects, not to servers
The Tallinn experts used the phrase in the same sense. They asked when civilian cyber infrastructure, itself a civilian object, is under a Party’s control. The majority answered territorially. All civilian cyber infrastructure in territory a Party controls falls under the rule. The minority took a narrower view. Military communications may pass through civilian servers over which a Party has no de facto control, and there the rule would not apply. All agreed that a system is under a Party’s control “if the party can dictate the operations of a civilian computer system” (Rule 121, para. 8). Given the feasibility caveat, the experts thought the split made only minor differences in practice (ibid.). Their question concerns civilian systems that need protection. Once military use turns a shared data centre into a military objective, it is no longer one of them (Rule 101, para. 1). What remains is the question whose civilians depend on it. None of the treatments considered for this post addresses that question for infrastructure hosted abroad. A wider survey of State positions may yet find one.
One facility, two circles of protection
Take first the domestic case. A Party’s forces use a data centre on its own territory, and its own hospitals and utilities depend on the same facility. Here everything lines up. The facility and the civilians who depend on it are all under the same Party’s control. The duty reaches all of them, and feasibility sets its limits. The Tallinn experts accept that segregation may not always be feasible. Where it is not, the Party must still take other measures, to the maximum extent feasible, for the civilians under its control (Rule 121, para. 7).
The cross-border case is different. I use the actors of post #31. States A and B are belligerents in an international armed conflict. Provider P runs a data centre in neutral State N. State A’s forces run military workloads there. The same facility hosts patient records for hospitals in State A and serves a water utility in State N. Whether State A may use P’s data centre at all is a question of neutrality law, which post #31 addressed. Here I assume that the use takes place. If the facility becomes a military objective, an attack on it endangers both groups. My conclusion is that State A’s Article 58 duty covers the hospitals in State A, because they are under its control, even though the servers are not. It does not cover the utility’s users in State N. Their protection rests on the attacker’s obligations and on the law of neutrality, not on State A’s Article 58 duty. The mirror case follows the same logic. A Party that hosts a facility at home that also serves users abroad owes Article 58 protection to its local dependants, not to the foreign users.
Two kinds of control
Article 58 thus works with two kinds of control. The first is legal and express. Control over civilians and civilian objects decides whom the duty protects. The second is factual and unwritten. A Party’s ability to dictate the operations of the facility it uses does not appear in the text, but it shapes what is feasible. A Party that runs its own data centre has more options at the facility itself. A Party that only rents capacity abroad has fewer. Its duty to its own dependants does not lapse for that reason. It shifts to the measures still within its reach. The test remains the familiar one of what is practically possible (1987 Commentary, para. 2245). The practitioner reading is that, for a force that relies on civilian cloud services, the test has two inputs: whose civilians are under the Party’s control, and what the Party can actually do about the facility.
One special rule might seem to reach further. Article 12(4) requires that medical units be sited, whenever possible, so that attacks against military objectives do not imperil their safety. A data centre that stores hospital records is not itself a medical unit. In my opinion, Article 12(4) therefore does not extend to it. The hospital’s dependency on the facility falls under Article 58(c) instead, and the control test of this section applies.
Conclusion
Article 58 does not forbid a force from using civilian cloud infrastructure. Nor does it settle whether such a force may share a facility with civilian users; that dispute remains open. What it does settle is the question this post set out to answer. The precautions against the effects of attacks protect the civilians and civilian objects under the using Party’s control. Control in this legal sense is tested for each of them, not for the servers. Where the facility and its dependants fall under the same Party’s control, the duty reaches them all and feasibility sets its limits. Where they fall under different control, as with services hosted abroad, the duty divides. It covers the Party’s own dependants, not the provider’s other civilian users.
Control over the infrastructure still matters, but in a different way: it shapes what is feasible, not whom the duty protects. For judge advocates advising a force that relies on services hosted abroad, two consequences follow. The duty does not end at the border of the host State. Nor does it travel with the servers to every civilian who uses them.