Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Systemic Cyber Incident Propagation Network Topology and Losses

Abstract. Systemic cyber loss arises when a digital incident reaches multiple economic interests through transmission, a common dependency, or the consequences of service interruption. Network topology constrains these mechanisms, yet the structure of a technological network alone cannot determine the distribution of insured losses. This monograph investigates the connection between directed and multilayer dependency graphs, incident propagation, business interruption, contractual indemnification, and insurance portfolio accumulation. It develops an analytical synthesis of actuarial network scholarship, official incident records, regulatory texts, court opinions, and insurance scenario research. A comparative examination of WannaCry, NotPetya, and the July 2024 CrowdStrike outage distinguishes malicious transmission, destructive activity, and nonmalicious common-source failure. Reproducible synthetic calculations isolate the effect of graph structure, loss dependence, business interruption duration, and insurance financial terms. The analysis establishes the conditions under which exposure concentration persists despite geographic or sectoral diversification and demonstrates why equal marginal expected losses can coexist with materially unequal aggregate tails. It identifies the limitations of endpoint counts, reported incident frequencies, disclosure data, and stress scenarios as estimators of insured loss. The conclusions concern mechanisms, measurement requirements, and the interpretation of conditional models; no empirical global catastrophe frequency or new named scientific theory is asserted.

Keywords: Systemic cyber risk; cyber insurance; network topology; dependency networks; contagion; common-source failure; software supply chain; business interruption; loss accumulation; insurance portfolio; tail risk; cyber catastrophe; reinsurance.

Introduction

The insurance significance of a cyber incident depends on the distribution of financial consequences across the interests insured. The interruption of a digital service can simultaneously prevent production, settlement, communication, logistics, and the delivery of public services. The same underlying technical disturbance can affect firms that have no direct contractual relationship with one another. An insurer may therefore hold a portfolio that appears heterogeneous under conventional classifications while concentrating exposure to a single software component, identity service, or infrastructure supplier.

The central research question is how the structure of digital and business dependencies translates into the severity and dependence of insurance losses. Its subsidiary questions concern the identification of plausible pathways, the measurement of conditional interruption, the translation of economic damage into contractually recoverable loss, and the aggregation of those recoveries at insurer and reinsurer levels. The distinction between a shared cause and transmission through a network is foundational. Confusing them produces different estimates of event footprint, conditional probability, intervention effectiveness, and the number of claims attributed to one occurrence.

This investigation extends the insurance science of dependency and accumulation rather than reproducing a general introduction to cyber coverage. The unit of analysis is an economically relevant interest associated with a specified policy, service, and time interval. Device failures and technical incidents enter the analysis as evidence about the mechanisms affecting that interest. Events can be correlated through service dependencies without a single infected machine transmitting malicious code to a second insured enterprise. Conversely, widespread infection may produce relatively little insured damage where affected functions are redundant or coverage does not attach.

The method combines critical reading of published research and institutional documents with comparative mechanism analysis and transparent, hypothetical calculations. The selected incidents serve as contrasting cases, not as a random sample or an estimator of catastrophe frequency. Official statistical surveys are interpreted according to their denominator and method of observation. Legal conclusions are bounded by the instrument, jurisdiction, date, and wording examined. The calculations are elementary consequences of established probability and network reasoning and can be independently reproduced from the assumptions stated. The economic and actuarial synthesis is the author's contribution. No proprietary portfolio, full machine-level infection telemetry, or independently estimated global loss distribution is available.

Conceptual and Historical Foundations

The object of systemic cyber insurance analysis

A systemic cyber incident is an event or connected set of events with consequences extending across multiple economically relevant units through shared exposure, technical transmission, or indirect dependency. Systemic significance is relative to the population and the outcome measured. A disturbance can be systemic for a regional hospital network and immaterial to a global insurer's aggregate capital, while another event can affect many insured businesses without materially impairing national economic output.

This relativity is essential for insurance science. A technical community defines the population through network connectivity and vulnerable components; a business defines it through functions, service continuity, and contractual relationships; an insurer defines it through the risks and financial interests it has accepted. An economy additionally includes households, public bodies, uninsured businesses, and infrastructure beyond the insurer's accounts. The boundaries need not coincide.

Awiszus et al. (2023) distinguish several sources of dependence in their actuarial analysis of cyber risks. The distinction provides a basis for separating loss variation specific to an insured from variation driven by a common environment or by interactions among exposed units. A portfolio can exhibit strong dependence through either a shared external factor or genuine propagation. In the latter case the state of one component changes the conditional risk of another; in the former a common cause changes both states. A real incident may combine the mechanisms.

For example, malware can enter through a shared software distribution channel and subsequently move through a firm's internal network. The delivery system supplies common-source dependence, while subsequent transmission produces conditional interaction. The resulting business interruption can then extend to customers whose technical systems remain intact. These three processes require three different observations: exposure to the initial event, transmission opportunities and actual state changes, and service or economic dependence.

Topology as a representation of dependencies

A graph G consists of a set of vertices V and a set of edges E. A vertex may represent a machine, identity system, software product, provider, business function, enterprise, or insured interest. An edge represents a specific relationship. Here an arrow from u to v means that a change at u can affect v under the mechanism under consideration. This convention is fixed throughout the study. Procurement databases often store the reverse relation, from customer to supplier; a graph imported from such a database requires orientation normalization.

Technical connectivity is commonly bidirectional at the communications level, whereas consequential dependence can be directional. A manufacturing firm might depend on a cloud-based order management service, although that service remains operational if this single customer stops production. A single graph with undifferentiated edges conflates communication, trust, control, service, and economic flow. A directed multilayer graph permits a node to participate in several relationships with different failure conditions.

A topology is not a complete description of a process. Weights, time, conditional activation, and recovery also matter. An edge can be active only during a software rollout, an authentication session, or a particular business function's operating hours. Two equally connected nodes may carry radically different amounts of critical traffic. A centrality ranking based on links alone therefore ranks a representation rather than a directly observed loss distribution.

Network density is the fraction of potential edges present under a defined node and edge convention. Degree records the number of incoming or outgoing connections. Directed reachability indicates whether any path connects a starting node to a target. Components and communities describe clusters under specified rules. Betweenness reflects paths that pass through a node. These quantities are useful diagnostics, provided they are tied to the dependency mechanism and observed graph boundary. A node central in a communication graph may be peripheral in a financial dependency graph.

Historical development of the problem

The contemporary cyber catastrophe problem developed through the interaction of network connectivity, standardized software, specialized infrastructure, and insurance coverage for digital losses. Concentration of functionality enabled economies of scale and operational benefits, while also creating exposures shared by organizations that previously relied on distinct local systems. The distribution of loss consequences increasingly came to depend on service architecture, not simply on the insured's local safeguards.

The 2017 WannaCry incident demonstrated the vulnerability of large organizations to widely distributed technical disruption. The UK parliamentary NHS cyberattack inquiry (2018) recorded shortcomings in preparedness and the difficulty of determining the full consequences. The resulting disruption implicated a public service system, not merely the machines on which malicious software ran. NotPetya in the same year illustrated destructive cyber activity with international economic consequences. The U.S. Department of Justice's 2020 charging announcement documents the allegations and selected victim losses associated with the episode.

The July 2024 CrowdStrike outage extended public attention to systemic disruption caused by a defective software update. The vendor's preliminary incident review described a limited distribution window and an affected Windows sensor population. The relevance of this incident is that widespread interruption can arise through routine trusted distribution. A cyber insurance model restricted to adversarial activity may omit such exposure even when policy language covers certain technology failures. A model should identify the insured peril independently of the technical cause.

These cases do not establish a time trend in event rates. They establish the diversity of mechanisms that require explicit treatment in a catastrophe scenario library. Historical incident categories remain useful only when the observed causes, affected populations, outcome metrics, and coverage provisions are consistently distinguished.

Primary Documents and Scope of Network Topology

Constructing the exposure graph

An exposure graph begins with a defined population and observation date. For underwriting and accumulation, the population comprises covered interests with policy periods, limits, retentions, coverage categories, and insured values. The graph must extend beyond the legal entities named on policies to services and suppliers whose failure could affect those interests. Identifiers should distinguish ultimate corporate groups, contracting subsidiaries, technology tenants, regional deployments, and particular product versions when those distinctions affect exposure.

The minimum practical structure is a dependency table recording the dependent entity, supplying component, affected function, geographic or technical deployment, type of dependency, criticality, alternative service, contractual relevance, observation date, source, and confidence. Each edge can be represented with a probability or interval of activation conditional on a specified event. Such a probability is a modeling parameter, not a substitute for the underlying evidence. Missing edges and uncertain dependencies require explicit coding.

The graph's boundary is an analytical decision with financial consequences. An insurer that observes only direct suppliers may fail to identify a common fourth-party service. A graph of ultimate providers may conceal regional segregation. A device-level graph can exaggerate the importance of an enterprise possessing many endpoints relative to one with a few commercially critical systems. Aggregation should follow the risk mechanism rather than the convenience of available identifiers.

Layers and projections

The technology layer records software, communications, shared control planes, authentication, storage and service delivery. The organizational layer records the firms, business functions, and service relationships whose operations depend on those technologies. The financial layer records insured interests, policy structures, group limits, and reinsurance arrangements. The layers connect through mappings, rather than through a single homogeneous adjacency matrix.

Consider three firms using the same provider. A graph projected onto firms may place edges between each pair because of their common supplier. Such a projection makes a triangle, although no firm transmits the incident directly to another. Introducing the provider as an explicit node preserves the mechanism: one initiating shock and three dependent responses. Estimating independent transmission probabilities on the projected triangle would count paths that have no demonstrated technical meaning. A projection is suitable when the purpose is to summarize shared exposure, but its edges must retain that interpretation.

The opposite problem occurs if service relationships are collapsed to an ultimate provider without retaining distinctions among products or availability zones. A shared corporate provider label can suggest universal simultaneous exposure even when dependencies are segregated. A provider failure scenario needs to specify the component and scope of failure before exposures are joined to the portfolio.

The multilayer representation should also distinguish failure of a component from loss of a service. Two alternative suppliers may provide resilience if alternatives are independently operable, sufficiently provisioned, and usable during recovery. If both alternatives require a shared identity or network service, their apparent independence is conditional on that third dependency. For a service requiring all essential inputs, interruption may occur when any one fails. For a genuinely substitutable service, interruption may require failure of all usable alternatives. These logical relationships are often more informative than counts of edges.

Security controls and evidence of topology

The NIST (2024) Cybersecurity Framework 2.0 organizes cybersecurity risk management outcomes. It can inform the collection of governance, identification, protection, detection, response, and recovery evidence. Its role in the present study is operational: control descriptions help determine which dependencies are relevant and what recovery or containment assumptions are plausible. The framework is not an empirical estimate of event probability or insured severity.

Underwriters should distinguish evidence that a control exists from evidence that it works under a specified scenario. An asset inventory may establish known exposure at a point in time; segmentation tests can inform the feasibility of movement; restoration exercises can support duration assumptions; service contracts and architectural records establish dependency. Each has a different evidentiary function. A questionnaire that reports an affirmative control without specifying scope and date cannot independently validate an entire graph.

In the European Union, the Digital Operational Resilience Act (Regulation (EU) 2022/2554, Articles 2, 29 and 64) supplies a regulatory example. Applicable from January 17, 2025, it establishes operational resilience requirements for covered financial entities, subject to its scope and exceptions. Article 29 addresses concentration at entity level, including substitutability, multiple arrangements with the same or closely related providers, and relevant subcontracting considerations. These provisions support the importance of tracking dependency concentration in financial institutions. They are not a rule for determining whether a policyholder's cyber loss is indemnifiable.

There is thus a useful correspondence, but not identity, between resilience mapping and insurer exposure mapping. The first asks whether the financial entity can continue its critical functions; the second asks how losses across insured entities accumulate. Insurers may need both perspectives because their own claims processing, communications, and settlement systems can depend on infrastructure also used by policyholders. Underwriting losses and the insurer's operational disruption should nevertheless be modeled and reported separately to avoid double counting.

Topological measures and their interpretive limits

The out-degree of a provider is an initial concentration indicator, but the number of its customers cannot establish financial accumulation without policy exposure and criticality. Weighted out-strength can be based on exposed limits or conditional interruption values. The interpretation changes with the weights: limits describe contractual capacity; conditional expected loss incorporates scenarios and assumptions about recovery. Gross limit sums are conservative capacity indicators in some circumstances, not automatic estimates of probable maximum loss.

Directed reachability identifies possible paths under a defined model, but not necessarily activated paths. A reachability measure is an upper envelope when every reachable node could fail and each edge is feasible, rather than a frequency estimate. Community structure can reveal pools of exposures with shared technology, but correlated risks may span communities through higher-level services. The relationship between graph centrality and insured loss is therefore an empirical question conditioned on node values, edge behavior, and coverage.

Spectral measures of a nonnegative transmission matrix can inform early-growth thresholds in specified linearized epidemic or branching approximations. The spectral radius is sensitive to weighted cycles and interaction structure. It is not a universal catastrophe threshold. A finite, acyclic directed chain has a nilpotent adjacency matrix with spectral radius zero and can still transmit a disturbance along its entire length. Common-source failure also produces widespread damage without requiring self-sustaining transmission. For insurance purposes, finite-event footprint and financial loss are as important as asymptotic growth.

Mechanisms and Evidence for Attack Propagation

Conditional transmission

A propagation model describes how the state of one component affects that of another during an event. A simple formulation assigns an edge-specific conditional probability p(u,v) that a disturbance at u reaches v within a defined period. The value can depend on vulnerability, credentials, segmentation, detection, control interventions, and the nature of the initiating event. Conditional independence of edge trials is a simplifying assumption requiring justification; the same patch status or security configuration may govern many edges simultaneously.

Time-dependent models can represent susceptible, affected, and recovered states. The utility of an epidemic analogy depends on the mechanism. Recovery from an interruption is not necessarily immunity to reinfection. A compromised endpoint may be isolated, restored, and later affected again. A provider failure can disable downstream services without a change in their infection state. A model should choose states and transitions according to their observable meaning.

Fahrenwaldt, Weber, and Weske (2018), in Pricing of cyber insurance contracts, describe an actuarial network approach combining an interacting Markov chain with conditional marked point processes. The publisher abstract supports the methodological connection between stochastic interaction and insurance claims. The present study relies only on that stated approach, not on unexamined numerical findings or a reproduced derivation. Hillairet et al. (2022) likewise investigate network-structured cyber contagion using a grouped representation of exposures; the available introduction and abstract establish the role of connectivity and group interactions in insurance-oriented modeling.

Both approaches motivate a distinction between the underlying state process and the observation or loss process. A state transition may be observable in telemetry but generate no immediate insurance claim. A claim may arise from downstream interruption rather than the insured's own compromise. In notation, a vector X(t) describes event-related component states at time t, while policy-level loss L(i) is a function of the relevant state histories, costs, and financial terms. The claim process cannot generally be identified by counting transitions alone.

Common-source exposure

A common-source event occurs when an upstream cause affects multiple downstream nodes through their exposure to the same source. There may be no direct onward transmission among the affected firms. A defective update, failure of an authentication service, or interruption of a shared hosting component can produce this pattern. A conditional model identifies a shock variable Z and exposures that respond to Z, with differences in version, timing, architecture, and recovery.

Even when responses are independent conditional on Z, their unconditional dependence may be substantial. For two losses, the law of total covariance yields

Cov(L(i), L(j)) = E[Cov(L(i), L(j) | Z)] + Cov(E[L(i) | Z], E[L(j) | Z]).

The first term captures dependence remaining within a defined event state; the second captures co-movement generated by variation in the common state. This decomposition is particularly valuable because an insurer can reduce direct technical transmission without eliminating dependence created by shared infrastructure. Conversely, a shared provider does not imply identical losses if conditional business responses differ.

Indirect economic propagation

Indirect propagation occurs when a firm's functioning depends on another entity's output, regardless of whether its own machines have failed. A manufacturer may be unable to receive orders, a retailer may lose payment functionality, or a warehouse may face interruption of dispatch instructions. The business effect depends on inventories, manual procedures, contract deadlines, alternative capacity, and the time necessary to activate substitutes.

Economic propagation can continue after the technical fault is corrected because queues and backlogs must be cleared. An interrupted supplier can transfer scarce recovery resources toward priority clients. A downstream business may partially recover before the supplier returns to full service. The relationship between cause and financial impact is therefore a time-dependent mapping. A technological recovery timestamp is not generally the end of the indemnity period, and the end of operational interruption is not necessarily the end of claim development.

A failure of a system that supports multiple sectors creates an exposure that can cross traditional underwriting classes. Interruption in healthcare, transport, professional services, finance, and retail need not be caused by direct interaction between those sectors. The shared dependency defines a cross-sector cluster. Sector diversification can reduce idiosyncratic losses without reducing this cluster exposure.

Reproducible topology experiment

Consider six nodes, with one initially affected node and five susceptible nodes. Compare two directed graphs, each with five edges. Graph A is a chain from the seed through five successive nodes. Graph B is an out-star in which the seed has a direct edge to each susceptible node. Assume each edge is attempted once when its source becomes affected, transmission succeeds independently with probability 0.5, and all affected nodes remain affected through the observation interval.

In the chain, the probability that the node at distance k is affected is 0.5 raised to the power k, because every preceding edge must succeed. The expected total affected nodes, including the seed, is 1 + 0.5 + 0.25 + 0.125 + 0.0625 + 0.03125 = 1.96875. In the star it is 1 + 5 × 0.5 = 3.5. The probability of affecting all six is 0.5 to the fifth power in both graphs, or 0.03125. This final equality follows because both graphs require success on every one of their five edges.

The distributions nevertheless differ. In the chain the number of affected downstream nodes has a truncated geometric distribution; in the star it has a binomial distribution. The chain can stop after one failed edge, preventing every further step. The star can still reach other targets after an edge failure. If each affected node creates a fixed $100,000 economic loss, conditional expected aggregate losses are $196,875 and $350,000 respectively. These are synthetic event-conditional outcomes. Neither graph estimates a real cyber event rate.

The experiment demonstrates that node count, edge count, and marginal edge probability do not uniquely determine expected aggregate loss. It also demonstrates that one tail statistic can coincide across distributions with different means and intermediate probabilities. Concentration analysis requires a distribution or a set of decision-relevant scenario quantities rather than one topological index.

WannaCry: propagation and service disruption

The parliamentary Cyber-attack on the NHS report examined WannaCry's consequences for England's health service and institutional preparedness. Its conclusions describe failures to act on warnings and weaknesses in the ability to assess the full cost and operational impact. This source supports the observation that consequences extended beyond infected devices and that contemporaneous cost measurement was incomplete. It does not establish the complete transmission graph or an insurance payment total.

For modeling, this case suggests three distinct data needs: the population of vulnerable or exposed systems, the population of operationally affected organizations and functions, and the economic consequences of disruption. The first two overlap but are not identical. A clinic that loses access to a shared service may experience interruption without being infected. An infected endpoint may be isolated without producing substantial interruption. A hospital's ability to redirect patients changes the distribution of outcomes across the health system.

Possible interventions have different scopes. Patching can reduce susceptibility; segmentation can constrain transmission; continuity procedures can reduce business or public-service impact. A model that treats each intervention as a reduction in a single undifferentiated loss probability obscures this causal structure. Estimating an intervention's financial effect requires understanding which of the distinct transitions it changes.

NotPetya: destructive activity and heterogeneous loss

The U.S. Department of Justice (2020) NotPetya charging account describes alleged state-linked destructive operations and reports nearly $1 billion in losses to three named victims. Those amounts are reported victim losses in a charging announcement. They are not a judicial determination of guilt, an industry loss total, insurer payments, or an estimate of worldwide economic damage. The source establishes historical evidence of severe multinational consequences while requiring careful interpretation of attribution and amounts.

For an insurer, the initiating distribution mechanism, subsequent propagation, and restoration burden are distinct parts of the causal account. Destructive activity can make restoration especially important to severity. The number of accessible nodes alone cannot quantify rebuild costs, data loss, operating downtime, or the possibility of maintaining partial service. Different businesses affected by the same initiating operation can therefore have substantially different severity distributions.

NotPetya also illustrates the significance of nonaffirmative cyber exposure in insurance lines that are not explicitly labeled cyber. The historical New Jersey appellate litigation discussed below concerned all-risk property policies. A catastrophe model must inventory relevant covered interests across potentially responsive lines, without presuming coverage from the technical categorization of an incident.

CrowdStrike: a distribution channel as a common cause

CrowdStrike's preliminary post-incident review identifies July 19, 2024, and a distribution interval of 04:09–05:27 UTC for the problematic content update. It identifies a population of eligible Windows sensors that received the content; Mac and Linux systems were not affected by that particular update failure. The later root cause analysis announcement confirms publication of further analysis. These are original vendor statements and should be interpreted with their date, scope, and corporate perspective.

Microsoft's Weston (2024) outage response statement estimated that 8.5 million Windows devices were affected, fewer than one percent of Windows machines. The number is an estimate of devices, not the number of insureds, disrupted businesses, claims, or loss dollars. Its financial relevance depends on the concentration of those devices in business-critical systems and the duration and nature of service degradation.

This episode is an instructive common-source scenario because the delivery of an erroneous update can affect many customers without malicious transmission among them. The exposure set is conditional on product version, platform, receipt time, and deployed architecture. A scenario built solely from the provider's customer list would overstate exposure by including customers outside the affected configuration. A scenario built solely from the endpoint count would fail to weigh business-critical functions and policy terms.

The cessation of problematic content distribution is also distinct from the recovery of affected systems. The original vendor review establishes the distribution window, not a uniform interruption duration for all customers. Recovery extends over different periods according to the mechanisms of access, repair, staffing, and continuity arrangements. This distinction should appear explicitly in event models and business interruption estimates.

Comparative mechanism assessment

WannaCry, NotPetya, and CrowdStrike share the potential for broad operational consequences, but they differ in their initiating mechanism, observed pathways, and appropriate exposure denominators. The parliamentary inquiry is concerned with a national public-service environment; the prosecution announcement identifies particular victim losses in a criminal case; the vendor and platform reports describe a software incident and affected endpoints. The sources cannot be combined into a homogeneous incident severity series.

The comparative conclusion is about mechanism, not ranking. A malicious propagation scenario requires information about conditional transmission. A common-source scenario requires identification of the exposed product and configuration population. An indirect business interruption scenario requires service dependency and recovery evidence. A combined event may require all three. Classification is useful when it guides evidence collection and contract evaluation.

Comparative Measurement of Financial Impact

Units and denominators

Measurement begins by defining the observational unit. Incidents, endpoints, organizations, policies, claims, occurrences, and affected business functions are different units. One incident can affect numerous organizations; one organization can hold multiple policies; one policy can receive several payments; multiple claims may be aggregated into one occurrence under applicable terms. A frequency measure defined over insured organizations cannot be compared without adjustment to one defined over devices or claim files.

The UK government's (2026) Cyber Security Breaches Survey reports that 43% of businesses and 28% of charities identified cyber breaches or attacks during the period studied. These are survey estimates of identified events for different populations. They are not claims frequency estimates and do not measure exposure to a specific common-source incident. The separate survey technical report documents the sampling and weighting procedures and the construction of cost questions. Identification and recall depend on awareness, reporting, and survey design; the cost of a most disruptive event cannot automatically be treated as the expected annual aggregate cost.

The U.S. Government Accountability Office (2021), in its cyber insurance market report, discusses limitations of historical loss data and coverage clarity, alongside evidence from broker-client insurance take-up. Such market reports can identify institutional changes and data problems, but broker-client populations are not national business registries. Take-up changes also alter the composition of the insured population. A rise in claim counts can result from coverage growth without a corresponding increase in underlying risk.

A dataset for loss modeling needs a time-varying denominator appropriate to the analysis. Policy-years can be useful for claim frequency; exposure intervals can be more appropriate where insured values or dependencies change within a policy period. The denominator should include observable zero-loss exposures rather than only claimants. Missing exposure information makes frequency and dependence difficult to identify even if the recorded claim amounts are accurate.

Ground-up, insured, and net loss

Ground-up economic loss is the loss borne by an economic interest before insurance compensation, under a defined accounting boundary. Policy-level insured loss is the covered portion after contractual financial terms. Gross insurer loss is the sum of obligations on the insurer's portfolio before ceded reinsurance, under a defined treatment of expenses. Net insurer loss incorporates recoveries and other applicable financing effects. These quantities need separate definitions in each model.

Ground-up losses may include direct restoration costs, loss of contribution from interrupted activity, incremental expenses, liability, and other measurable consequences. They can also involve public costs and changes in welfare that are not business accounting losses. Summing supplier revenue losses and customers' lost production can count the same transactions more than once when the goal is an economy-wide welfare or output measure. Summing distinct contractual indemnities is a different operation, governed by each insured interest and the limits of duplicate recovery.

A useful business interruption calculation starts from the counterfactual operating result during the period of interruption. Lost revenue alone omits avoided variable costs. Lost contribution or another policy-defined business income measure, adjusted for mitigation and relevant extra expense, can better reflect interruption damage. The counterfactual requires seasonality, order timing, demand, inventories, and partial capacity where those factors matter. It should distinguish permanently lost sales from deferred sales.

Duration and insurance transformation

Consider an illustrative insured business earning a contribution of $100,000 per normal operating day. Assume total interruption lasting four days, no deferred sales or relevant avoided expenses beyond the specified contribution calculation, a one-day waiting period for interruption, $50,000 covered restoration expense, a $25,000 monetary retention applied once to the combined covered amount, and a $250,000 policy limit. Under these assumed terms the covered interruption before retention is $300,000. The combined amount is $350,000; after retention it is $325,000, and the limit reduces payment to $250,000.

If the interruption lasts two days under the same assumptions, covered interruption is $100,000, and the payment is $125,000 after adding restoration and subtracting retention. Ground-up consequences and insured amounts change differently. Under the four-day example an additional short period of interruption may increase economic damage without increasing payment because the limit is exhausted.

The calculation is illustrative rather than a particular policy interpretation. Some forms treat waiting periods as temporal deductibles; others specify different triggers and expense structures. Retentions and limits may apply separately to different coverages. The transformation must be coded according to the actual wording and ordering of financial provisions, and the policyholder's loss must be established before indemnification is calculated.

Dependence and aggregate distribution

For individual insured losses L(i), aggregate gross loss A = Σ L(i). Its expected value is the sum of individual expectations, regardless of dependence when expectations exist. Its variance is the sum of individual variances plus twice the sum of pairwise covariances. Dependence can change capital requirements and tail probabilities without changing the marginal expected claim cost.

Consider ten insureds, each with $250,000 of hypothetical covered loss per affected event, a $50,000 retention, and a $200,000 payment limit. Assume no other loss during the modeled period. In portfolio I, each insured independently incurs the $200,000 payment with probability 0.02. In portfolio II, a common provider shock occurs with probability 0.02 and affects all ten, each generating the same $200,000 payment. Otherwise none is affected. Both portfolios have identical marginal loss distributions.

Quantity Independent portfolio Common-shock portfolio
Expected aggregate loss $40,000 $40,000
Aggregate standard deviation $88543.77 $280,000.00
Probability of at least $1,000,000 loss 7.41464037e-7 0.02
99% value at risk $400,000 $2,000,000

The independent portfolio's affected count follows a binomial distribution with parameters 10 and 0.02. The probability of at least five affected insureds is the sum over k from five through ten of C(10,k) × 0.02^k × 0.98^(10−k). The common-shock portfolio pays $2 million in the shock state and zero otherwise. Value at risk is defined here as the smallest loss whose cumulative probability is at least 0.99. Atoms in the distribution are retained, rather than smoothed through an arbitrary continuous approximation.

The comparison isolates dependence. It does not represent a calibrated insurer portfolio or an actual cyber catastrophe return period. In a real portfolio, dependence can also change marginal frequencies and severities through varying controls, recovery, and selection. A model must not infer that dependence affects only variance in all circumstances. The fixed-marginal comparison shows why a collection of independently priced risks can nonetheless generate a materially different capital exposure when insureds share a common source.

Quantile-based capital measures also require care where there are rare, severe events. A single quantile can fail to reflect the scale of consequences above it. Tail expectation measures, full loss exceedance curves, stress scenarios, and sensitivity analysis provide additional information. Each should specify whether the distribution is annual aggregate, occurrence-based, conditional on an event, or conditional on an event family.

Conditional catastrophe scenarios

The 2018 Lloyd's and AIR Worldwide Cloud Down scenario report modeled a three-to-six-day outage of a top-three cloud provider. Its stated ranges included $6.9–14.7 billion in ground-up loss and $1.5–2.8 billion in insured affirmative cyber loss. These are conditional estimates from a specified historical modeling exercise, not actual reported payments, an updated forecast, or the expected annual loss of the cloud sector.

The scenario is valuable because it explicitly separates economic damage from insured loss and considers concentration around major providers. Its results cannot be transferred mechanically to a contemporary portfolio. Provider infrastructure, insurance penetration, coverage terms, business dependencies, prices, and recovery techniques may change. A new analysis should restate the event assumptions, exposure date, dependency mapping, and contract population before applying a model.

A stress scenario establishes what a portfolio could lose under specified conditions. It does not by itself establish the probability that those conditions occur. To estimate annual exceedance probabilities, a model needs event occurrence assumptions or a defensible stochastic mechanism. Where frequency estimation is weak, a set of conditional loss distributions and explicitly chosen stress conditions is more transparent than a single unqualified return-period number.

Quantifying uncertainty

Uncertainty arises in the event initiation process, the exposure graph, propagation transitions, service interruption, financial severity, and policy interpretation. It is useful to distinguish variation within a stipulated model from uncertainty about the model itself. A simulation that repeats edge trials estimates conditional stochastic variation; it does not automatically capture unobserved suppliers or an incorrect event definition.

A practical sensitivity analysis changes one or more consequential assumptions: the fraction of customers sharing the relevant configuration, probability of effective segmentation, distribution of recovery duration, availability of alternatives, attachable coverage, and policy limits. Alternative graph completions represent uncertainty about missing dependencies. Scenario results can be reported as intervals across these assumptions, with a record of which combinations are feasible. An interval is only as informative as the rationale for its endpoints.

Calibration is especially challenging when historical events are few and heterogeneous. Fitting one model to a small set of prominent catastrophes can confuse model sensitivity with evidence. For example, the largest known outage is not automatically an extreme quantile of the current insured portfolio. Changes in dependency architecture and underwriting terms can invalidate stationarity. Backtests should separate event footprint, duration, gross economic loss, and insured financial loss where observations permit.

Integrated Scientific Synthesis and Research Agenda

The relationship between topology and insured tail risk

The evidence and calculations support a conditional proposition: network topology influences systemic cyber loss through the propagation and dependency mechanisms operating on that topology, while economic and contractual structures determine the distribution of financial consequences. A graph supplies possible relations; an event supplies activation; affected enterprises translate disruption into loss; insurance contracts determine which losses are transferred. Estimation requires evidence at each stage.

This sequence helps identify the location of model failure. An incorrect provider mapping is an exposure error. An implausible transmission probability is a mechanism error. A uniform recovery duration for heterogeneous enterprises is a consequence-model error. A misapplied exclusion or limit is a financial transformation error. Correcting one stage does not automatically correct the others. The value of a modular model lies in making these errors diagnosable.

The synthetic graph comparison establishes that equal numbers of nodes and edges can produce unequal conditional expected footprints. The synthetic insurance comparison establishes that identical marginal risks can produce unequal aggregate tails. Together they show why simple exposure counts and individual expected losses cannot fully characterize systemic loss. This is an analytical synthesis of established methods, not a proposed new theory or universal empirical law.

Validation across stages

Model validation can be organized around separate observable outcomes. The first is exposure validation: does the mapped population match documented providers, components, and configurations at the date of the event? The second is propagation validation: do predicted paths and affected populations match reliable technical observations? The third is economic validation: do duration and functional impairment correspond to observed operating consequences? The fourth is insurance validation: do modeled amounts and claim categories correspond to adjudicated or appropriately developed financial observations?

A failure to observe a particular outcome is informative about the limits of validation. Publicly reported endpoint counts may help validate a broad technical footprint while leaving financial validation entirely open. A signed claim settlement may establish a particular insurance outcome without reconstructing the technical path. Modelers should report validation by stage, rather than describing an entire catastrophe model as empirically validated from one type of observation.

The empirical cases in this monograph validate the existence and distinction of several mechanisms through original institutional evidence. They do not supply a joint dataset sufficient to estimate all model parameters. Their use is consequently qualitative and comparative. The numerical results are reproducible demonstrations under explicitly stipulated assumptions, not fitted statistics.

Data requirements and selection effects

An improved research dataset would connect anonymized policy exposure information, dated dependency records, event classes, affected services, interruption durations, recovery evidence, and developed claim amounts. Secure data governance would be necessary to protect sensitive infrastructure maps and confidential insurance information. Public summaries could report aggregated concentration and uncertainty measures without disclosing operationally sensitive details.

Selection effects deserve particular attention. Insured firms differ from uninsured firms in size, activity, resources, controls, and propensity to report. Increased underwriting scrutiny can change the insured population over time. A dataset assembled from paid claims selects on contractual eligibility and reporting in addition to technical incident occurrence. These factors make it difficult to identify the causal effect of topology from a simple regression of paid losses on provider exposure.

A more defensible study would establish a population before an initiating event, identify exposure configurations for affected and unaffected units, and observe relevant outcomes under consistent definitions. Observational comparisons require adjustment for confounding and recognition that dependency choice itself may reflect firm characteristics. A quasi-experimental design may sometimes exploit an event whose exposure is tied to a particular rollout or configuration, but causal interpretation requires evidence that the identifying assumptions hold. This monograph supplies no such estimated causal coefficient.

Research priorities

Further empirical research should compare service-level graph reconstruction with provider-level aggregation and determine how much financial exposure is concealed by coarse representations. Specific attention should go to substitute capacity and recovery: a nominal alternative service may become unusable under simultaneous demand, while modular architectures may permit partial operations. The appropriate outcome is measurable continuity or financial consequence, not merely the existence of an alternative edge.

A second research priority is the construction of representative exposure denominators. Current public surveys, regulatory reports, and incident accounts contribute useful information, but their units and sampling procedures differ. Linking them without adjustment creates misleading precision. Collaborative data standards should preserve observation windows, event taxonomy, coverage scope, development age, and the denominator corresponding to each outcome.

A third priority is the validation of accumulated insured losses across multiple lines and contract forms. Changes in policy wording can alter the relationship between economic loss and insurance response without changing technical risk. Scenario libraries should therefore version financial assumptions alongside exposure and hazard assumptions. A catastrophe model constructed for historical coverage may require substantial adaptation when applied to a contemporary underwriting portfolio.

A fourth priority concerns international comparability. Technical dependencies can be global, while contracts, litigation, supervisory frameworks, accounting definitions, insurance penetration, and public support differ by jurisdiction. An international model can share technical scenarios across portfolios while applying locally appropriate legal and financial modules. Comparability is strongest when scenario assumptions are common and financial transformations are transparent.

Scope and limitations of the present research

The study uses accessible original documents and scholarly publication records. The Fahrenwaldt et al. source is limited to the publisher abstract and metadata; the Hillairet et al. source is limited to accessible abstract, introductory and section material. The relevant DORA and Merck passages were examined as extracted official PDF text; page-image retrieval was unavailable for those documents. The historical Lloyd's materials were examined as PDF text and page images where retrieved. The CrowdStrike review is explicitly preliminary, and the later announcement is used only for the fact and context of the RCA release.

Source access does not justify claims beyond the material examined. No claim of a systematic census of all published network models, a complete review of every insurance wording, current post-2023 Merck procedural status, or the current status of all national cyber catastrophe policies is made. The legal materials are illustrative; the financial calculations are synthetic. No proprietary insurance data, incident-level matched portfolio dataset, or independently validated empirical event-frequency distribution was available.

The findings are consequently suited to methodological interpretation, research design, and the explanation of systemic insurance loss. Application to a particular portfolio requires its exposure information, policy wording, validated event assumptions, and appropriate professional and institutional review.

Conclusion

Systemic cyber incident propagation is consequential for insurance because the distribution of economic interests and their dependencies differs from the distribution of devices or direct technical connections. Shared technology can introduce substantial dependence into an otherwise diversified portfolio. Propagation can extend the initiating footprint, and economic dependencies can spread financial interruption beyond directly affected technical systems. Insurance terms then determine the recoverable portion and group it according to contractual conditions.

The three historical incident types examined establish the relevance of malicious propagation, destructive activity, and accidental common-source failure. They also demonstrate the inadequacy of a universal observational unit for cyber catastrophe measurement. An endpoint estimate, an institutional disruption report, and selected victim losses are valuable evidence for different questions. Their proper use depends on the populations, mechanisms, and outcomes each records.

The quantitative illustrations show how topology changes a conditional event footprint and how dependence changes aggregate loss despite identical marginal insurance risks. These conclusions imply that underwriting and capital assessment should preserve shared dependencies, contractual differences, and uncertainty over event initiation and recovery. The science of cyber insurance becomes more informative when these relationships remain visible from the original incident record through the modeled indemnity and ultimate portfolio loss.

The principal research requirement is the integrated observation of technology exposure, time-dependent service consequences, and developed insurance loss. Until matched evidence is sufficiently representative, transparent conditional modeling and carefully delimited comparative analysis provide a stronger foundation than unqualified global loss or return-period estimates.

References

  • Awiszus, K., Knispel, T., Penner, I., Svindland, G., Voß, A., & Weber, S. (2023). Modeling and pricing cyber insurance: Idiosyncratic, systematic, and systemic risks. European Actuarial Journal, 13, 1–53. doi:10.1007/s13385-023-00341-9. Publication record.

  • Fahrenwaldt, M. A., Weber, S., & Weske, K. (2018). Pricing of cyber insurance contracts in a network model. ASTIN Bulletin, 48(3), 1175–1218. doi:10.1017/asb.2018.23. Publication record.

  • Hillairet, C., Lopez, O., d’Oultremont, L., & Spoorenberg, B. (2022). Cyber-contagion model with network structure applied to insurance. Insurance: Mathematics and Economics, 107, 88–101. doi:10.1016/j.insmatheco.2022.08.002. Publication record.

  • National Institute of Standards and Technology. (2024, February 26). The NIST Cybersecurity Framework (CSF) 2.0. NIST CSWP 29. doi:10.6028/NIST.CSWP.29. Publication record.

  • European Parliament & Council of the European Union. (2022, December 14). Regulation (EU) 2022/2554 on digital operational resilience for the financial sector. Official Journal of the European Union, L 333, December 27, 2022. Articles 2, 29, 64. Publication record.

  • House of Commons Committee of Public Accounts. (2018). Cyber-attack on the NHS. Thirty-second Report of Session 2017–19, HC 787. Conclusions and recommendations. Publication record.

  • U.S. Department of Justice, U.S. Attorney’s Office for the Western District of Pennsylvania. (2020, October 19). Six Russian GRU officers charged in connection with worldwide deployment of destructive malware and other disruptive actions in cyberspace. Publication record.

  • CrowdStrike. (2024, July 24; updated July 25). Preliminary Post Incident Review (PIR): Content Configuration Update Impacting the Falcon Sensor and the Windows Operating System (BSOD). Publication record.

  • CrowdStrike. (2024, August 6). Channel File 291 incident: Root cause analysis is available. Publication record.

  • Weston, D. (2024, July 20). Helping our customers through the CrowdStrike outage. Microsoft Official Blog. Publication record.

  • U.S. Government Accountability Office. (2021, May 20). Cyber Insurance: Insurers and Policyholders Face Challenges in an Evolving Market. GAO-21-477. Publication record.

  • U.S. Government Accountability Office. (2022). Cyber Insurance: Action Needed to Assess Potential Federal Response to Catastrophic Attacks. GAO-22-104256. Publication record.

  • Merck & Co., Inc. v. ACE American Insurance Co. (2023, May 1). Nos. A-1879-21, A-1882-21. Superior Court of New Jersey, Appellate Division. Publication record.

  • Lloyd’s. (2024, May 14). State backed cyber attack wordings. Market Bulletin Y5433. Publication record.

  • Lloyd’s & AIR Worldwide. (2018). Cloud Down: Impacts on the US Economy. Emerging Risk Report, pp. 6–7. Publication record.

  • Department for Science, Innovation and Technology & Home Office. (2026). Cyber Security Breaches Survey 2025/2026. Official statistics. Publication record.

  • Department for Science, Innovation and Technology & Home Office. (2026). Cyber Security Breaches Survey 2025/2026: Technical Report. Publication record.