The war exclusion says nothing about the digital, and that is the whole problem. It was written to carve out of cover a risk deemed non-mutualizable, belligerence between states, at a time when an act of war was recognizable by its means: troops, borders, a declaration. It still sits in property policies written long before cyber was treated as a distinct risk. Applied to a computer attack conducted or sponsored by a state, it asks a text designed for tanks to classify a piece of software, and the question is not whether that can be argued but who bears the uncertainty.
The market learned this lesson on a real case. The 2017 NotPetya attack was attributed by the American and British governments to Russian military intelligence, and several insureds sought to trigger conventional property policies that did not mention cyber, a phenomenon known as silent cyber. Insurers raised the inherited war exclusion. Two American cases came out of it, and they did not end the same way. Mondelez against Zurich was settled in 2022, during trial, without a judgment on the merits. Merck, by contrast, was decided: in 2023 the New Jersey appellate court held that the inherited exclusion, drafted for traditional warfare, did not reach this attack and found for the insured, before a settlement reached while the case was before the state Supreme Court. The lesson is not that a war exclusion never operates in cyber. It is that its inherited drafting leaves the legal uncertainty with the insurer, since an exclusion is a defense and the burden of proof follows the party raising it.
The market's answer was to rewrite the clause rather than litigate it. In August 2022, through Market Bulletin Y5381, Lloyd's of London required a cyber war exclusion in every standalone cyber policy written by its syndicates from March 31, 2023. In November 2021 the Lloyd's Market Association had published four model clauses, LMA 5564 through LMA 5567, which differ in how much cover they keep, from the total exclusion of LMA 5564 to the widest write-backs of LMA 5567. In January 2023 it replaced them with eight texts, an A and a B version of each, and the only difference between A and B lies precisely where everything is decided: the treatment of attribution.
That is where two rules meet. The A versions contain an agreement on how a cyber operation is attributed to a state: the insured and the insurer consider the objectively reasonable evidence available to them, which may include an official attribution by the government of the state where the affected system is located, and the burden of proof stays with the insurer. They are the versions that satisfy Bulletin Y5381 for standalone cyber policies. The B versions contain no such agreement, and attribution is then proved under the law governing the contract, with no agreed method. On the same fact, with the same insured and the same damage, these two wordings do not open the same debate. The first gives a particular place to a political act performed by a third party; the second leaves the whole question to the judge or the arbitrator. Neither is neutral, and choosing between them is not a drafting detail, it is the allocation of legal risk.
The costliest asymmetry is temporal. A public attribution is a diplomatic act, with its own calendar, decided for reasons foreign to the claim: it may land a few weeks after the facts, several years later, or never. A clause that gives weight to that act inherits that calendar. An insurer may then raise an attribution that occurred after the file was closed, and a settlement signed on the strength of an absence of attribution is weakened. Conversely, some states deliberately sponsor criminal groups while maintaining ambiguity about their status, which makes attribution structurally unavailable and leaves indefinitely open a question on which payment depends.
Then comes the insurance tower, where the gap turns invisible. Excess layers follow the primary, but only as far as their own exclusions allow. A tower assembled across three renewals, with several carriers, can perfectly well carry an A version on the primary and a B version on a high layer, or an LMA 5567 below and an LMA 5564 above. The same fact is then covered below and excluded above, and the insured discovers the hole at the moment of need. A placement summary listing capacities and premiums never shows this: it would have to align the wordings, which few summaries do.
Finally, one and the same attack often presents itself under two contracts at once, and they do not carry the same clause. A property policy predating the cyber affirmation movement carries an inherited war exclusion and no cyber exclusion; a recent standalone cyber policy carries a modern cyber war clause. The claim is then examined twice, on two different definitions of war, and nothing guarantees the two insurers will retain the same classification of the first cause. The order in which the insured presents its claim stops being a procedural question and becomes a substantive one.
A French industrial group operates a subsidiary in a country neighboring an armed conflict. On April 8, a wiper spreads through the update of a local accounting package and destroys the group's ERP, including in France. Business interruption is assessed at 18 million euros. Twelve million are presented under a property policy written in 2019 on the French market and tacitly renewed since, which does not mention cyber and carries an inherited war exclusion. Six million are presented under a standalone cyber policy written in 2024, carrying a clause from the LMA 5564 to 5567 family in its A version, which has the parties weigh the objectively reasonable evidence, including an official attribution by the government of the state where the affected system is located. On September 2, two foreign governments publicly attribute the attack to a state intelligence service. Both insurers raise the war exclusion.
The September 2 attribution appears to settle both files at once, and that is precisely the reasoning that proves expensive, because the two contracts do not give it the same role. Under the 2024 cyber policy, attribution is at the heart of the exclusion without being a formal condition of it: the clause has the objectively reasonable evidence weighed, and the official attribution it cites is that of the government of the state where the affected system is located, here France for the group's ERP. A joint statement by two foreign governments is weighty evidence, not the evidence the clause designates, and the burden of proof stays with the insurer. The discussion therefore bears on the quality of the evidence and on its date, not on the reality of the attack. Under the 2019 property policy, one question comes before war: is the destruction of an ERP physical damage within the meaning of the policy? Under most wordings, erased data is not, and without physical damage there is no business interruption to exclude. If the policy clears that first threshold, attribution plays no contractual role at all: the clause speaks of war and hostilities, notions the litigation arising from NotPetya found poorly fitted to a cyberattack, and since an exclusion is a defense, the burden sits with the insurer. Attribution there becomes an argument of fact, not a condition satisfied. The reflex of presenting the claim first under the broadest policy therefore runs against the file: it is the 12 million under the 2019 policy, the one whose exclusion is weakest, that carry the stake, and that is where the classification of the first cause is decided. Finally the date matters as much as the fact: if the clause weighs an official attribution without saying when it must exist, an attribution arising after a settlement can be raised, and a settlement signed in July on the strength of an absence of attribution is weakened. What this file teaches is not that a more favorable clause would be needed, but that a tower whose layers do not carry the same definition of attribution is not a tower, it is two covers stacked on each other.
- 01A cyber war exclusion that relies on attribution ties the outcome of the claim to a third party's diplomatic calendar, which can run slower than settlement.
- 02Clauses LMA 5564 to 5567 differ in the cover they keep; their A versions agree on a method of attribution, the B versions contain none: same fact, two debates.
- 03The war exclusion is a defense: on twentieth-century drafting, the burden of proof stays with the insurer, as the litigation arising from NotPetya showed.
- 04A tower whose layers do not carry the same definition of attribution opens a hole that the capacity summary never shows.