RESEARCH DOSSIER
Cyber Incident Frequency Trends and Exposure Denominators
Abstract. Cyber incident frequency is the occurrence of a specified class of digital events per defined quantity of exposure over a stated observation interval. Its measurement requires a reproducible event boundary, a population at risk, a time reference, and an account of how incidents become observable. This monograph examines incident prevalence, recurrent event intensity, asset-time rates, revenue-scaled counts, and the role of policy exposure as a distinct insurance quantity. It integrates official enterprise surveys from the United Kingdom, Australia, Canada, and the European Union with regulatory reporting mechanisms, incident-response standards, threat intelligence corpora, United States complaint statistics, insurance market reporting, and original empirical cyber-risk research. A documented fall or increase in reported counts may reflect changing hazards, expanding digital exposure, improved detection, revised reporting obligations, changing composition, or several mechanisms jointly. Formal count-process, observation, standardization, and sensitivity analyses clarify the conditions under which trends are identifiable. Worked examples demonstrate reversal of crude trends by exposure growth and composition, partial identification under uncertain detection, distinctions between recurrent counts and prevalence, and the restricted interpretation of insurance-policy ratios. The conclusions establish a disciplined interpretation of available evidence rather than a universal global incident rate or new named theory. No proprietary microdata, causal intervention estimates, or independently calibrated global hazard series are claimed.
Keywords: Cyber incident frequency; exposure denominator; enterprise prevalence; asset-years; detection probability; cyber insurance; reporting bias; event attribution; risk standardization; longitudinal comparability; count models; observation process; digital exposure; insurance statistics; cyber risk science.
Introduction
A statement that cyber incidents have increased is scientifically incomplete until its event class, observation system, population and period are established. An incident count describes a corpus. A prevalence statistic describes the fraction of eligible units affected during a defined window. An incidence rate describes recurring events relative to exposure-time. A conditional distribution describes the characteristics of events that reached a dataset. Each answers a legitimate question, and each can change in a different direction during the same historical period.
The practical importance of these distinctions has expanded with the digitization of commerce. An enterprise may add thousands of devices, migrate workloads to shared infrastructure, increase the number of customer accounts and establish dependencies on external identity or payment services while remaining one enterprise in the business register. If its incident count grows, an analyst interested in enterprise continuity, an engineer interested in endpoint compromise and an insurer interested in the probability of a covered loss face different statistical tasks. The unit that appears in the denominator establishes the question being answered.
This Research Dossier belongs to the Science Cyber Insurance. It investigates the occurrence and observation of cyber incidents before the distinct transformation of incidents into insurance claims. It treats the insurance interface insofar as underwriting, portfolio design and interpretation of market statistics depend on incident-frequency evidence. The separate scientific problem of joint claim frequency and severity is outside the principal scope.
The research proceeds by comparative documentary analysis and formal analytical illustration. The official survey waves are examined at the level of their published estimates, eligible frames and comparison restrictions. Regulatory documents establish why administrative reporting selects events. The historical empirical literature tests the proposition that differently assembled loss and breach records need not support identical trends. Mathematical examples are constructed solely to expose identification conditions; they are not observations or estimates. Source availability varies: the complete relevant HTML sections of several official and academic records were accessible, while selected regulatory and threat-landscape records were available through indexed primary text, and the NAIC report through text extraction. These limits constrain the strength of the corresponding interpretations. A public global longitudinal census with matched digital asset-time denominators was not obtained in this study.
Three questions organize the analysis. Which observation and counting conventions establish cyber incident frequency? Which denominator corresponds to each substantive risk question? Under what conditions do apparent temporal and international differences support a conclusion about underlying risk? The answers require treating the numerator and denominator as a single measurement design.
Conceptual and Historical Foundations
The event and the population at risk
A cyber incident is operationally defined here as a realized episode meeting a predetermined cybersecurity event criterion within a specified entity, system or service boundary. The criterion may concern unauthorized access, confidentiality compromise, integrity impairment, disruption of availability or another documented departure from security requirements. The exact scope must be adapted to the scientific question. In particular, the inclusion of attempted attacks, fraud, accidental failures and near misses must be declared before counting.
The incident-response perspective locates this measurement within preparation, detection, assessment, containment and recovery. Nelson et al. (2025), in NIST SP 800-61 Revision 3, integrate incident response into broader cyber risk management. The publication provides an institutional basis for treating observation capabilities as part of organizational security. A monitoring program that discovers more events may be improving its capacity to identify them. A statistical record that ignores this relationship can interpret institutional learning as worsening security.
The relevant historical transformation is from localized computing incidents toward organizational exposure distributed across cloud services, identities, software components, communications and interdependent firms. The legal entity still provides a useful economic unit, but its digital boundary is increasingly permeable. Responsibility for preventing an event, first detecting it, suffering consequences and communicating it may be distributed among several parties. An organization may be economically exposed to an incident that never compromises its own devices.
The event definition should distinguish the generating episode from its affected units. Suppose a compromise at a software supplier affects 60 customers. A generating-event dataset may contain one event, an enterprise-impact dataset as many as 60 customer events, and an endpoint dataset thousands of affected machines. All three representations can be valid if their identifiers and counting conventions remain distinct. A time series that moves from supplier-level counting to victim-level counting will exhibit an artificial discontinuity if the change is unrecorded.
A comparable distinction separates an attack attempt, a confirmed incident, and an incident with material consequences. In an incident funnel, an attempted credential theft may be blocked, result in successful compromise, or lead to business interruption, notification expenses or liability. Event records describing attempts help characterize threat pressure, while confirmed compromises support a different hazard measure. Incident impact is a distinct mark attached to the realized episode. Insurance is concerned with financial consequences within the coverage contract, subject to reporting and settlement processes. A well-specified research database preserves these stages and their timestamps.
Incidence, prevalence and repeat events
Let N(i,t) denote the count of qualifying events attributable to entity i within time interval t. Let Y(i,t) be one when the count is positive and zero otherwise. The total event count is the sum of N(i,t), whereas the affected-unit count is the sum of Y(i,t). For a complete frame of M entities with equal annual observation, enterprise prevalence is P = sum(Y)/M. The mean annual event count is F = sum(N)/M, with F at least P whenever counts and the affected-unit definition agree. These are distinct functions of the same underlying count distribution.
An affected-entity survey can measure P without estimating F. Even perfect measurement of prevalence does not determine the number of repeated incidents. With 100 organizations and 40 affected, a corpus could contain 40 distinct organization-event pairs or thousands. The same prevalence is compatible with very different recurrence burden. Conversely, a higher average count may be concentrated in already affected enterprises and leave prevalence nearly constant.
Under the restrictive homogeneous Poisson assumption with a common integrated annual intensity lambda, P = 1 − exp(−lambda). If P = 0.40, the implied lambda is −log(0.60), or approximately 0.511 events per organization-year. That inference fails for a heterogeneous population unless the distribution of intensities is specified. In a mixed population, P = 1 − E[exp(−Lambda)]. By Jensen's inequality, the implied homogeneous rate −log(1 − P) is at most E[Lambda] under the mixed-Poisson assumptions. This inequality is a mathematical diagnostic, not an unconditional conversion rule for survey data.
The saturation of prevalence is important. Moving from one to ten events in an affected enterprise has no effect on Y. Where many firms experience at least one low-level attempt, the binary indicator may be insensitive to large changes in attempts per firm. A different threshold, such as an incident involving verified unauthorized access, can generate a more discriminating prevalence statistic for a different substantive question. Threshold changes must remain explicit.
Historical evidence and competing signals
The archival literature shows why provenance matters. Edwards, Hofmeyr, and Forrest (2016), in Hype and heavy tails, analyzed a historical US publicized-breach corpus and found no modeled trend in frequency or size within their studied interval. Their count fit and heavy-tailed size analysis explain why unusually large annual changes can arise without an estimated trend. Their result refers to the assembled corpus, definitions and historical window; it does not establish constancy for all later cyber events.
Romanosky (2016) used incident rates by industry in a broader economic analysis, connecting observed events with the number of firms in an industry. This is an explicit movement beyond raw counts. Its interpretation still depends on the correspondence between the database coverage and the population statistics. Commercial event collection may be more complete for some sectors or firm sizes than others. The useful denominator therefore reduces one source of ambiguity while leaving observation selection to be assessed.
A later analysis by Shevchenko et al. (2023), The nature of losses from cyber-related events, examined a much wider set of cyber-related event categories, reporting increases in recorded frequency over 2008–2016 and substantial sector and event-type heterogeneity. Its mainly US enterprise representation and distinct commercial database differ from the privacy-breach chronology used by Edwards and colleagues. These findings can coexist without implying a scientific contradiction. They describe different event taxonomies, observability mechanisms, intervals and corpora. Demonstrating a genuine contradiction would require holding these features sufficiently constant.
The historical literature consequently supports a bounded conclusion. Cyber event archives contain evidence of occurrence patterns, but the scale and direction of a population hazard trend require denominators, exposure histories and observation-process information that are seldom all present in a single corpus.
Frequency Measurement and Observation Processes
From occurrence to a recorded event
A statistical observation can be represented as the result of sequential eligibility and information processes. An event occurs at a time S. It is detected at D, classified at C, reported at R and ingested into a dataset at A. These times may coincide for a conspicuous outage but can differ considerably for a covert intrusion. The occurrence date belongs to the hazard process; the detection and reporting dates belong to the observation process. Administrative arrivals should be modeled as arrivals when occurrence dates are unknown, rather than retroactively treated as a complete occurrence series.
Let E be an exposure measured in a specified unit-time and lambda an average occurrence intensity per unit of exposure. In a simple independent thinning model, the expected observed count equals E × lambda × q, where q is the probability that an eligible event becomes a recorded observation. This encompasses detection, recording, reporting and dataset inclusion, with their conditional structure absorbed into q. Under perfect unit attribution and common q, dividing the observed count by E estimates the product lambda q. Only with independent information on q, or a defensible stability assumption, does the observed rate identify an underlying intensity or its trend.
False positives, duplicates, and classification errors require an extended account. If expected erroneous additions equal b, then expected observations are E × lambda × q + b. When multiple sensors report the same event, deduplication rules affect b. If classification errors move events across categories, a matrix of class sensitivities and misclassifications replaces the scalar observation probability. The scientifically appropriate level of precision is governed by the evidence available about that measurement process.
Consider a monitoring system that records 80 eligible incidents among a constant population in the first year and 100 in the second. If detection improved from 0.60 to 0.90 and all remaining assumptions of independent thinning hold, the corresponding expected-count estimates are 133.3 and 111.1 after division by q. The observed count increased 25%, while the illustrative corrected estimate decreased by about 16.7%. These values are hypothetical. They demonstrate why stable detection, not merely stable exposure, is necessary for reading an observed count change as a hazard change.
Unit attribution and event reconciliation
Event construction is a measurement task prior to modeling. A record should contain a stable incident identifier, organization and system identifiers, generating-event or campaign linkage where available, affected-service identifiers, occurrence and detection time ranges, an event type and classification confidence, source provenance, an impact indicator, and a record-status field. It should preserve revised records as a versioned history rather than counting each update as a new event.
Repeated credential attempts may form one campaign, numerous attempts or several successful compromises depending on the chosen unit. A defensible time-window rule should be grounded in operational continuity and the triggering event, not selected afterwards to flatten a trend. A supplier incident that is later attributed to a common vulnerability can be linked to a generating event without erasing the customer-level observations needed to measure business exposure.
The analyst should preserve positive, negative and unknown states. An organization with no incidents detected in a fully monitored year differs from one whose monitoring data are missing. An organization that left the frame midway through a year differs from a continuously observed organization. An unknown occurrence date should remain unknown or be encoded as an interval rather than replaced silently by the publication date. These rules make subsequent denominator estimates auditable.
Reporting institutions as observation mechanisms
The statutory and administrative record samples events according to the scope of each institution. Under the European Union's NIS 2 Directive (European Parliament and Council, 2022, Article 23), significant incidents of covered entities enter a notification framework. Under DORA's major incident reporting (European Parliament and Council, 2022, Article 19), financial entities report major ICT-related incidents to the competent authority. These legal texts establish classes of reportable events and regulated populations. National implementation, technical criteria and applicable sectoral rules must be examined separately before constructing jurisdiction-specific time series. The present discussion concerns the reporting mechanism and does not adjudicate an individual institution's compliance obligations.
The SEC's (2023) cybersecurity disclosure compliance guide explains that domestic Exchange Act registrants ordinarily report a material incident on Form 8-K within four business days of determining materiality, while foreign private issuers operate under different disclosure provisions. The relevant selection is therefore public-company status and materiality, rather than every attempted attack. A filing date is generally later than occurrence and does not directly measure the beginning of an attack. An apparent upward step after a reporting requirement takes effect may be an observation-regime change even when the underlying process is unchanged.
Voluntary administrative reporting creates another frame. The FBI (2026) IC3 complaint summary reported 1,008,597 complaints received in 2025, compared with 859,532 in 2024. The increase is approximately 17.34%. This is evidence about volume entering the complaint center, not an estimate of the change in the fraction of US businesses compromised. Complaints may concern individuals, enterprises, diverse fraud and cyber-enabled crime categories; the numerator has no matched universal population of exposed organizations. Public awareness, reporting access, crime mix and actual victimization can all affect volume.
The correct research question for a reporting agency may nonetheless be arrival burden: how many cases require classification or referral per investigator-time? The denominator, in that case, is investigative capacity rather than firms or assets. A shift in scientific purpose changes the legitimate rate. The same administrative data can support operational workload measurement without constituting a population cyber incidence census.
Contributor corpora and case composition
Verizon's (2026) DBIR data collection overview describes contributions from law enforcement, forensics, legal services, insurers, sharing groups and its own caseload, with an in-scope interval running from November 1, 2024 to October 31, 2025. This kind of contributor network is valuable for investigating attack mechanisms and distributions among observed cases. A denominator consisting of classified breaches can support a conditional proportion such as the share of observed breaches attributed to a given mechanism, when classification and missingness are properly handled.
The total number of covered enterprises and their exposure-time across all contributors, including noncases, is a separate question. A contributor's addition may expand observed cases without changing the underlying hazard. A complete case corpus from a single investigator network may therefore remain incomplete as a population sampling design. Comparisons across editions need contributor continuity, taxonomy continuity and exposure-to-inclusion information.
ENISA's Threat Landscape 2025 publication record describes 4,875 analyzed incidents across July 2024 through June 2025. Its Threat Landscape 2026 record uses January through December 2025. The windows overlap for six months and are not consecutive annual intervals. A comparison of their headline totals as annual growth would mix overlapping observations, possible changes in collection and dataset coverage. The publication metadata are used here exclusively for the stated corpus and interval boundaries; direct retrieval of the complete ENISA texts was unavailable.
A cross-source catalog should avoid double counting episodes reported by several agencies or vendors. Record linkage can match timestamps, victims and mechanism attributes, but those attributes may be redacted or too ambiguous. An analyst should quantify unresolved linkage and preserve alternative match thresholds. An exact match policy can miss genuine duplicates; an aggressive probabilistic policy can merge separate incidents. A sensitivity range is scientifically preferable to an unjustified claim of a deduplicated global event total.
Exposure Units and Denominator Construction
Enterprise-years and dynamic entry
For an organization-level event process, define exposure E as the sum over enterprises of the amount of time each enterprise satisfies eligibility and is observed. An enterprise observed for an entire year contributes one enterprise-year; an enterprise entering at midyear contributes roughly one-half, assuming observation begins then. Event counts must be restricted to corresponding intervals. The rate r = N/E has units of events per enterprise-year.
This formulation handles births, deaths, mergers and gaps in observation. When two firms merge, the treatment depends on whether the scientific entity is the legal corporate record, a consistent economic group or a system boundary. A common legal identifier history can document changes while preserving separate operational identifiers. A fixed-panel design answers a question about continuously observed firms; an open-cohort design answers a question about the changing population. Both are necessary to separate survivor effects from wider market change.
An annual snapshot count of firms can be used as an approximation to enterprise-years where intra-year movement is small or the reporting convention requires it. It should be labeled as such. The approximation becomes particularly problematic in fast-growing digital industries or periods of business closures. Counts from surviving firms alone may omit firms that exited following adverse outcomes, producing informative attrition.
Enterprise prevalence instead uses eligible firms in a well-defined sampling frame and common recall window. A weighted estimate has the form sum(w_i Y_i)/sum(w_i). The resulting point estimate is not simply the unweighted proportion of respondents. Valid uncertainty calculations need the survey design, including weights, clustering, stratification and any finite population adjustments. The interviewer response or sampling contact is not itself the population unit.
Asset-time and identities
Asset-years, device-days, workload-hours, account-months and internet-facing-service-days represent technical opportunity for selected classes of events. The choice should follow the causal exposure mechanism. Credential compromise plausibly relates to active identities and access opportunities, while endpoint compromise may be studied in device-time. An application vulnerability may demand application-time with a relevant component or accessible interface.
Let A_i(u) count eligible active assets for organization i at instant u. Its integrated exposure during observation is the integral of A_i(u) over time. Adding devices halfway through a quarter should not give each new device a full quarter of exposure. An inventory snapshot can be used to approximate the integral only when the inventory and activity are sufficiently stable, or a documented interpolation is defensible. Ephemeral cloud containers and transient workloads make snapshot methods particularly susceptible to error.
The asset denominator must use a stable counting grain. Virtual machines, container instances, services, physical devices and software components have complex relationships. Counting both hosts and their virtual instances as independent interchangeable assets can inflate exposure. Identity denominators similarly distinguish human users, privileged accounts, service principals, inactive accounts and machine identities. A single human may control multiple accounts; a service account may be used by many processes. The appropriate denominator follows the event's mechanism and scope.
The numerator must correspond to the denominator. One enterprise incident affecting 500 devices divided by device-years yields enterprise episodes per device-year, a descriptive scaled rate. It is not the probability of device compromise. To measure the latter, count distinct affected devices or device-events and specify repeated compromise, incident grouping and recurrence. Both rates may help: enterprise episodes reflect response workload, and device-events reflect technical spread. A change in the number of affected devices per enterprise episode can otherwise be mistaken for an increase in event generation.
Asset-time is also insufficient where opportunity depends on active use or contact volume. A parked device and an internet-facing authentication service are not equally exposed simply because each exists. Subdivision by network exposure, privilege, patch state, control deployment and workload is often more defensible than a single synthetic asset score. Missing inventory should be separately identified because undercounting the denominator inflates rates even with a stable numerator.
Transactions, records and revenue
A fraudulent transfer or payment deception may be studied per initiated payment, per authenticated session or per total transaction value. These rates measure distinct quantities. Incident count per million transactions measures frequency normalized by activity; lost dollars per million dollars transferred measures a loss ratio relative to throughput. A database containing only known fraudulent payments and no legitimate transaction volumes cannot estimate transaction incidence.
Personal-data breach analysis may use the number of records held, accounts served or sensitive records accessible. Records exposed in detected breaches are an outcome measure rather than a complete exposure denominator. A study of compromised records per million records held requires a time-matched inventory and a rule for duplicates, backups and shared records. Combining the inventory of data copies and the number of unique individuals affected obscures the unit of analysis.
Annual revenue provides a scale measure familiar in business and underwriting. The rate of organization-events per billion dollars of revenue can be a useful economic intensity measure. Revenue is neither a direct count of opportunities for a technical attack nor automatically proportional to hazard. A retailer with high turnover, a software firm with a large identity population and a capital-intensive manufacturer can have the same revenue with materially different digital structures.
Suppose Firm A earns USD 100 million annually and records 10 incidents; Firm B earns USD 1 billion and records 20. Firm B has twice as many recorded incidents per firm-year, while Firm A has 100 per billion dollars of revenue versus 20 for Firm B. There is no contradiction. Enterprise response burden and economic scaling have different denominators. A revenue normalization should specify the accounting period, currency conversion, inflation basis and treatment of consolidated group revenue. Deflation can be necessary for comparisons in constant purchasing-power terms, while nominal insured revenue may be relevant to contractual financial exposure.
Revenue is also potentially endogenous: an outage can lower realized revenue during the same year. Normalizing loss-inducing incidents by postincident revenue creates a feedback in the denominator. A lagged or baseline revenue measure can reduce this particular mechanical relationship when the question concerns ex ante scale. The selection must be documented; it does not alone establish causal identification.
Insured exposure and contractual translation
A policy-year is one year of covered exposure under a specified contract. Earned policy-years sum actual coverage durations and are preferable for a claim-occurrence rate when policies enter and leave during the observation period. Policies in force at a stated date are a stock. An annual count of claims divided by that stock may be a useful published-market ratio, but it differs from an exposure-year rate unless equivalence is justified.
The enterprise-to-policy relationship is many-to-many. A corporate group may have primary, excess and endorsement coverage under different insurers. Some contracts cover liability, interruption or cybercrime exposures in distinct ways. One technological event can generate several notices, affected insured entities or insurance files, while numerous other incidents generate no covered loss. A count of insurance claims depends on the definition, contract boundary, retentions, waiting periods, reporting behavior and allocation of affected interests.
In general, let K_e be the number of claim records arising from event e within a selected portfolio. The total number of claim records attributable to events is the sum of K_e. K_e can be zero, one or greater than one. Expected claim count depends on frequency and the expected multiplicity conditional on event characteristics and coverage. Without matched enterprise incident records, policy information and claim linkage, the conversion from claims to incident occurrence is underidentified.
Coverage denominators require treatment of retroactive dates, claims-made or other reporting provisions, cancellation, endorsements, limits and changing contract definitions as applicable. The contractual exposure population must be clearly fixed before interpreting an incidence rate. The present monograph develops the distinction between enterprise and insurance statistical universes; it does not interpret a concrete policy form that was not retrieved.
Shared technology and exposure at dependencies
A shared provider introduces multiple legitimate exposure levels. Provider-years measure generating-event exposure. Customer-provider relationship-years measure the risk of customer impact. Workload-hours hosted by a provider measure a technical quantity of dependence. A single common event creates dependence among customer impacts, so these exposure quantities do not imply independent event outcomes.
The central data need is an explicitly dated mapping from enterprises to services and providers, with shared dependencies identified where feasible. A portfolio with many firms can remain highly concentrated in one technology dependency. Counting enterprise-years without measuring shared dependency obscures aggregation risk, even if the resulting average enterprise event rate is well estimated. An average can characterize marginal frequency while providing little information about simultaneous losses.
An illustrative provider event impacts 80 of 1,000 customers. Provider event count is one; affected-customer proportion is 8% for that event. If the provider was observed for one year, the generating-event rate in that one-provider observation is one per provider-year. The 80 customer impacts are dependent. A confidence interval treating them as independent events would understate the uncertainty of the generating event process. With more providers and repeated independent or appropriately modeled common shocks, both the generating frequency and conditional propagation can be studied.
Longitudinal Trends and International Comparability
United Kingdom enterprise evidence
The UK 2025 breaches survey provides a comparable pair of business prevalence estimates: 50% in its 2024 comparison and 43% in the 2025 release. The 2025/2026 survey reports 43% again. The observations concern identified attacks or breaches within a recall window, rather than all latent successful compromises. The reported fall is seven percentage points, equivalent to a 14% proportional decline from the rounded 50% starting value. The two latest point estimates are equal at the reported precision; this is not proof that the underlying unrounded probabilities are identical.
| Survey comparison wave | Business prevalence | Interpretation |
|---|---|---|
| 2023/2024 | 50% | Comparable overall-question baseline |
| 2024/2025 | 43% | Decline in reported prevalence |
| 2025/2026 | 43% | Same rounded point estimate |
The questionnaire change identified in the 2025 release restricts comparisons with waves preceding the 2024 question design. Consequently the three-wave comparison should be maintained as a bounded series. A researcher can examine other earlier metrics if their wording remains comparable, but should not insert results from a discontinued question into this particular sequence.
The most recent publication distinguishes enterprise sizes. Reported prevalence is 42% among microbusinesses, 46% among small businesses, 65% among medium businesses and 69% among large businesses. This pattern is consistent with differences in opportunity, operations and detection, and does not identify the separate contribution of each factor. Organization-level prevalence tends to grow with opportunities even if comparable technical components have stable risk. More capable detection at large firms provides another plausible explanation.
The survey technical report documents disproportionate sampling and iterative weighting of business and charity respondents, with attention to anomalous numerical answers. Weighting allows the sample to approximate the target population across observed frame characteristics. Residual nonresponse or unequal identification probabilities can remain. Sampling uncertainty and measurement uncertainty are therefore different dimensions of precision.
The 2025/2026 survey reports that among businesses with an identified breach or attack, 29% experienced one at least weekly. If the conditional and marginal survey estimates share the stated eligible population, a rough product of rounded values is 0.43 × 0.29 = 0.1247, or approximately 12.5% of all in-scope businesses. This derived quantity is an illustrative approximation to the joint proportion, not an official reweighted estimate or a precise number of weekly events. Exact joint estimates would require appropriate microdata weights and uncertainty calculations.
The same logic applies to source classifications. A category share among affected enterprises uses a denominator smaller than all enterprises. If the affected population and conditional prevalence are both weighted consistently, their product approximates all-enterprise prevalence for the specified category. Because a business can experience more than one attack type, the category shares need not total 100%. The analyst should preserve the frame and conditionality of every published percentage.
These results answer an important substantive question: within a bounded and broadly comparable UK survey design, the recent overall observed prevalence fell and then remained at the same rounded value. They do not establish the trajectory of per-asset incidents, repeat-event counts or financial losses. Each additional inference demands its corresponding measurement.
Australia, Canada and the European Union
The Australian Bureau of Statistics (2023) Characteristics of Australian Business publication reports 22% of businesses experiencing an internet security incident in 2021–22 and 8% in 2019–20, while explicitly warning that the questions changed. The later instrument described eight types of incident and allowed multiple answers. The numerical difference of fourteen percentage points is a difference between reported measurements, not a defensible estimate of an underlying prevalence increase. A bridge survey administering both formulations to comparable samples would provide better evidence of the break.
Statistics Canada's 2023 cybercrime release and 2025 cybercrime release report that approximately 16% of Canadian businesses were impacted by cybersecurity incidents in each reference year. The later indexed release places the 2019 estimate at 21%. The 2023 survey metadata defines a business-register frame across specified industries and excludes enterprises with fewer than ten employees. The relevant outcomes are incidents impacting enterprises. These figures are treated as institutionally reported, rounded prevalence estimates; direct opening of the two release pages did not succeed, although indexed primary text was retrieved. The 2025 methodological details were not independently recovered in a frozen version during this research.
Eurostat's Digitalisation in Europe (2026 edition) presents approximately 22% of businesses experiencing ICT security incidents with consequences in its 2024 results. The associated enterprise survey scope (Eurostat, 2024) identifies selected NACE sections and an employment or self-employment threshold of at least ten persons. The reference calendar year for incidents must be reconciled with the survey year before joining the estimate to annual event series; the interactive overview's year label alone is insufficient to establish that alignment.
| Source and reported period | Published measure | Frame and comparability issue |
|---|---|---|
| UK 2025/2026 wave | 43% of businesses | Identified attacks or breaches; includes small microbusinesses |
| Australia 2021–22 | 22% of businesses | Revised incident question; break with earlier wave |
| Canada 2025 | 16% impacted | Impact threshold; business survey frame; 2025 version not frozen in retrieval |
| EU 2024 survey results | About 22% | ICT incidents with consequences; specified enterprise frame; year alignment requires care |
These statistics should not be interpreted as a league table of national cybersecurity performance. The UK measure can include a broader class of identified attack experiences, the Canadian measure restricts reporting to impacts, and the European frame uses its own industry and employment conventions. Canada and the EU exclude or restrict parts of the smallest-enterprise population through eligibility definitions. Differences in firm composition, survey period, question wording and detection may explain part of any numerical gap.
International comparisons are more defensible if limited to common sector and size cells, with compatible incident definitions and reference periods. Even then, different laws, awareness and security capabilities create reporting variation. International scope is scientifically achieved through explicit comparability conditions rather than by forcing disparate numbers into a common ranking.
Insurance market comparison and denominator growth
The NAIC (2026) cyber insurance market report provides a sharp example of why exposure scaling matters. Its extracted 2025 reporting-year material gives 5,843,175 policies in force versus 4,368,614 for 2024, together with approximately 20% growth in reported claims to 56,323. It reports substantial endorsement representation in the policy count. The available text and transcribed figure values were inspected; visual page screenshots could not be fetched. The interpretation is therefore restricted to explicitly accessible extracted figures.
The stock of policies grew by (5,843,175 / 4,368,614 − 1), approximately 33.75%. If counts and policy stocks had strictly matching scopes, the relative change of a crude reported-claim-to-policy-stock ratio would be approximately 1.20/1.3375 − 1, or −10.3%. This calculation uses a rounded claims growth figure. The 2025 ratio itself is 56,323/5,843,175 × 100, approximately 0.964 reported claims per hundred policies in force. Neither calculation is a calibrated incident hazard, a unique insured-enterprise rate or an earned policy-year frequency. Their interpretation also depends on claim timing and the alignment of market reporting categories.
This is an insurance illustration of a general denominator problem. Absolute count growth can coexist with a decline in a chosen normalized count. The stock, composition and contract design of the insured population changed. Aggregate membership is therefore endogenous to market pricing, underwriting and product choices. A separate actuarial study using matched earned exposures, consistent claim definitions and occurrence-year data would be needed to estimate changing claim frequency. Enterprise incident trends require a different population.
Conditions for a comparable time series
A valid longitudinal comparison holds its estimand constant or explicitly adjusts for changes. The minimum metadata consist of event definition, population eligibility, dates, observation mechanism, unit of count, duplicate policy, inclusion criteria and denominator construction. A break in one of these elements is a candidate measurement discontinuity rather than immediate evidence of a stochastic regime change.
Suppose a researcher receives incident counts from different security providers. The earliest provider observed 200 firms, the later provider 500, and the providers used different incident classifications. A denominator adjustment from counts to firm-years would address expansion only if eligible firm-time were measured. It would not reconcile the event classification. A harmonized classification must be applied retrospectively to records, or the analyst must compare only stable categories and report missingness.
Data availability over time may also produce asymmetric retrospective coverage. Older events can be added through archival disclosures, while recent periods remain incomplete due to detection delays. A database extracted today is a vintage: it contains observations available under the current collection process. Freezing successive snapshots permits an analysis of revisions, and specifying a common development age reduces recent-period truncation. The hazard estimate and the forecast of unobserved reports should remain separately identifiable.
In an international longitudinal study, changing classification standards can be treated as segmented series, sometimes with bridging estimates. A dummy variable for a known reporting regime change can describe a discontinuity; it cannot establish whether the underlying hazard also changed at that time. Independent data on reporting uptake or a stable control series strengthens attribution. The design should avoid describing a fitted coefficient as a causal effect solely because it has a statistically significant value.
Evidence Validation and Sensitivity Analysis
Exposure-adjusted counts and composition reversal
All calculations in this chapter use hypothetical data unless a source is explicitly identified. Consider two annual observation periods with 100 and 120 confirmed incident episodes, and aggregate exposures of 1,000 and 1,500 enterprise-years. The raw count increases by 20%, whereas the observed rates are 0.10 and 0.08 episodes per enterprise-year, a 20% reduction. With constant classification and observation probability, the second statistic is more relevant to the rate of enterprise occurrence.
| Hypothetical period | Recorded events | Enterprise-years | Events per 100 enterprise-years |
|---|---|---|---|
| A | 100 | 1,000 | 10 |
| B | 120 | 1,500 | 8 |
The denominator growth is 50%. Decomposing total events as rate times exposure gives a count ratio of 1.20 = 0.80 × 1.50. The arithmetic identifies the contribution of exposure scaling within the chosen observational system; it does not show that underlying hazard changed if observation probability also changed.
Composition adds another layer. Suppose two size groups have deliberately distinct exposure-specific rates. A portfolio moves toward large enterprises, while each group's rate decreases.
| Hypothetical group | A exposure | A events | A rate per 100 | B exposure | B events | B rate per 100 |
|---|---|---|---|---|---|---|
| Smaller enterprises | 900 | 45 | 5 | 600 | 24 | 4 |
| Larger enterprises | 100 | 20 | 20 | 400 | 72 | 18 |
| Combined | 1,000 | 65 | 6.5 | 1,000 | 96 | 9.6 |
Both strata became safer by their own observed rates, while the combined rate rose about 47.7%. Direct standardization to period A's 90% and 10% exposure weights gives B a rate of 0.9 × 4 + 0.1 × 18 = 5.4 per hundred, compared with A's 6.5. The standardized decline is 16.9%. This is an example of composition-driven reversal (Simpson's paradox). It strengthens the case for publishing both crude and standardized series.
An exact two-term decomposition with A weights distinguishes the within-stratum component from change in weights evaluated at B rates. In units per hundred, the first equals 5.4 − 6.5 = −1.1, and the second equals 9.6 − 5.4 = 4.2; their sum equals the observed increase of 3.1. Another reference weighting changes attribution between components without changing the total difference. Researchers should state the decomposition convention rather than presenting its terms as invariant causal effects.
For fixed E values and independent homogeneous Poisson events, the variance of N/E is lambda/E. In the first example the approximate log rate ratio standard error is sqrt(1/100 + 1/120), approximately 0.1354. The rate ratio is 0.8; a conventional 95% Wald interval is exp(log(0.8) ± 1.96 × 0.1354), approximately 0.614 to 1.043. Sampling fluctuation is compatible with no rate change under those assumptions. This interval is purely illustrative and would be inappropriate if event dependence, uncertain exposure or unequal observation were substantial.
Partial identification through observation probabilities
Suppose observed rates per enterprise-year are 0.10 and 0.08, and credible independent audits indicate q_A lies in [0.50, 0.80] and q_B in [0.60, 0.90]. Under the simple no-false-positive observation model, A's underlying expected rate lies in [0.125, 0.20], and B's in approximately [0.0889, 0.1333]. The true rate ratio equals the observed ratio multiplied by q_A/q_B, so the range is approximately [0.4444, 1.0667]. The observed decline alone does not resolve the sign of the underlying change across this sensitivity range.
These are logical plug-in bounds conditional on the assumed ranges, not confidence or credible intervals. A sampling model for observed counts and a distribution or sampling design for q would be necessary for probabilistic uncertainty. Bounds can also be tightened with repeated calibration, known sentinel systems, or independent record linkage when the required assumptions hold. If false positives are nonnegligible, their frequency must be bounded or estimated in the extended observation model.
Detection can vary systematically by incident type, severity, business size and governance. A single q is then a weighted average whose weights may themselves change. Estimates at a sufficiently stable stratification level are more informative. For example, detection of service interruption can be nearly certain from uptime metrics while credential compromise may remain invisible until much later. Combining their observed counts into one aggregate with an assumed common q obscures both processes.
A capture–recapture illustration clarifies the strengths and risks of overlapping records. If two lists contain 80 and 50 incidents and 20 appear in both, the simple independence estimate of total detectable incidents is 80 × 50 / 20 = 200. The two lists contain 110 distinct observed incidents. Under independently homogeneous capture, the missing count would be estimated at 90. Positive dependence in capture—such as severe incidents more likely to appear in both lists—can bias the estimate downward. More than two lists, covariates, explicit linkage uncertainty and sensitivity analyses may help, but the assumptions of population closure and matching remain essential. The numerical example is didactic and is not derived from real cyber datasets.
Count-process models
A standard model for an observed count N_it with exposure E_it is a Poisson regression with log(E_it) as an offset: log(mu_it) = log(E_it) + alpha + beta t + gamma'X_it. A time coefficient beta corresponds to a multiplicative observed-rate change conditional on the covariates, exposure definition and model. If an estimated detection probability q_it is treated as known, log(q_it) can also be included as an offset under the thinning model, with its estimation uncertainty propagated separately.
An offset fixes proportional scaling with the exposure quantity. This is a strong assumption for technical inventories or revenue. Where the count is expected to scale nonlinearly, the model may estimate the coefficient on log(E) rather than fixing it to one. A coefficient below one would imply sublinear association in that specification; it would not prove that adding assets causes safer operation. The estimand changes when E enters as a modeled explanatory variable rather than a strict offset.
Poisson conditional mean and variance equality often fails when organizations differ in latent exposure, when attacks occur in campaigns or when episodes trigger repeat attempts. The negative binomial variance form Var(N|X) = mu + phi mu squared permits overdispersion when phi is positive. The historical publicized-breach frequency study (Edwards et al., 2016) illustrates empirical use of a negative binomial count fit. Model choice for an enterprise cohort requires testing overdispersion, residual patterns and predictive performance in that cohort; a fit in a public disclosure archive is not transferable automatically.
Zero inflation or hurdle models distinguish a propensity to observe at least one event from recurrence among positive observations. Such models can be useful for heterogeneous controls or observation regimes, but a structural-zero component must have a defensible interpretation. Zeros may indicate actual event absence, limits of detection, inactivity, missing data, or a narrow event definition. Statistical mixtures alone cannot tell which explanation is true.
Longitudinal enterprise panels can include fixed or random effects to account for persistent heterogeneity and time effects for common shocks. Fixed effects identify within-entity associations over changing covariates, subject to model and sample assumptions. Random effects introduce a distribution for unobserved variation and can support prediction for new entities. Firms entering and exiting observation require explicit treatment. Exposure and controls measured after an incident may be consequences of that incident; temporal ordering and lagged measures are indispensable when assessing prevention.
Time-varying intensity models can represent changing threats or recurrent events. An intensity conditional on history may increase after a compromise, reflect a campaign, or decrease after containment. A fitted self-exciting process may predict clustering, but simultaneous common shocks and recording batches can produce similar patterns. Process selection requires external evidence and temporal resolution sufficient to distinguish these mechanisms.
A single standard error based on independent observations is generally insufficient for shared-provider shocks. Cluster bootstrap or robust methods can be considered when there are enough meaningful independent clusters. In portfolios with several overlapping provider exposures, the choice of cluster structure is substantive. For rare catastrophic events, scenario analyses may be more transparent than a precise empirical tail estimate. The frequency of ordinary incidents and the frequency of common-shock events should not be conflated.
Validation of survey estimates and event cohorts
A prospective validation procedure begins with the estimand and event dictionary. It next reconciles the exposure frame independently of the event list. The event ledger is deduplicated under declared incident grouping rules. Follow-up time is calculated for each eligible unit. The analyst reports the crude rate, size and sector strata, a standardized estimate with fixed weights, the rate of missing data, and the observation-process assumptions. Only then should a trend model be selected.
Survey-derived prevalence can be validated against alternative measures of detection or independent benchmarking but cannot be treated as complete microdata without access to the underlying sample design. Appropriate confidence intervals require the actual survey standard errors or a design-consistent calculation. The point estimates in the UK, Canadian, Australian and European publications remain published descriptive estimates, not harmonized re-estimates. No p-values have been calculated for differences between their populations.
A validation audit of a case database should quantify the share of reports lacking occurrence dates and the share with only disclosure dates. Where lag distributions are estimable, occurrence cohorts can be examined at a fixed reporting development age. Unknown lag parameters should be represented by sensitivity scenarios. If reporting delays differ by incident class, aggregation of the lag distribution before correction can create distortions.
Exposure records deserve the same scrutiny as event records. Asset inventories may drift because discovery tools changed. A system migration may generate duplicate identities, ghost accounts, or short-lived workloads. Annual revenue classifications may be restated. Insurance policy counts may include endorsements alongside dedicated products. Revisions must be independently versioned so a rate can be reproduced from the exact numerator and denominator versions used in its calculation.
A falsification-oriented analysis asks whether the claimed trend appears in a measure that should not be affected by a specific observation change. If new monitoring mainly records attempted credential abuse, compare the documented change with a separately stable severe-outage series where appropriate. A divergence can be informative about ascertainment, although attack mechanisms may also evolve differently. A negative-control measure must have credible scientific grounds; choosing it after observing results creates additional selection risks.
Scientific interpretation of evidence strength
The strongest inference supported here is documentary and mathematical. Several official statistical series define affected populations and disclose comparison limitations, while event corpora and administrative reports establish their own data collection constructs. The algebra shows that count ratios cannot identify rate ratios absent exposure data, and observed rate ratios cannot identify hazard changes absent appropriate observation assumptions. These conclusions follow independently of a particular dataset.
The empirical inference is narrower. The UK recent wave comparison is directly documented within the official survey design. The Canadian and Australian releases identify distinct impact and question scopes. Cross-country figures remain descriptive because no harmonized microdata were acquired. The ENISA and legal source comparisons rely on limited retrievable primary text in this session; detailed classification and numerical assertions beyond that evidence are excluded. The NAIC figure calculations use the extracted text, with failed visual retrieval documented. The academic articles are interpreted within the data and fitted models the authors report, without claiming independent replication from proprietary or archived microdata.
These limitations prevent an absolute global trend estimate, not a scientific investigation of the measurement problem. The research contribution is the explicit correspondence between each supported empirical signal and its legitimate estimand, including the demonstration of how a change in denominator, population composition or ascertainment can reverse an apparent conclusion.
Integrated Scientific Synthesis and Research Agenda
What the evidence permits
The comparison establishes that the trend question must be expressed with its unit. An enterprise prevalence trend measures change in the probability that an eligible organization identifies at least one event in the specified window. A recurrent enterprise incidence trend measures change in event burden per enterprise-time. Asset-level incidence measures events at a technical grain per corresponding asset-time. A complaint series measures administrative intake. A claims ratio measures a coverage-related outcome in a selected portfolio. Each is a valid scientific object when its construction is transparent.
The direction of one series imposes few restrictions on the direction of another without further assumptions. For example, improved monitoring may increase identified attacks while improved controls reduce successful compromise. Expanding digital exposure may increase total events while reducing per-asset risk. A decline in organization prevalence may coexist with greater concentration of repeated events among affected firms. Changes in financial impact may occur through severity, propagation and recovery cost independently of count trends. A portfolio's reported claims can rise while enterprise attack prevalence elsewhere falls.
The analytical synthesis therefore favors parallel reporting of complementary indicators. Occurrence, detection, prevalence, recurrence, impact, and administrative reporting can be recorded with separate denominators. A compact scientific publication should state the estimand in the title, table headings or notes so that the reader can reconstruct what numerator and denominator produce the result.
Application to underwriting and risk engineering
An insurance underwriter can use the documented incident environment as contextual evidence for selecting covariates, scenario types and security questions. Without data from noncases and exposure-time, case distributions should not be used directly as unconditional loss probabilities. Enterprise-level and asset-level data offer different insights: the first measures overall organizational burden, while the second can inform assessment of specific security mechanisms.
Selection into insurance is an additional institutional process. Firms may buy coverage because they face high digital dependence, because lenders or contracts require it, or because of prior losses. Insurers can change eligibility and pricing as information develops. A portfolio therefore differs from an economy-wide register. Comparisons with insured cohorts should condition on observable selection variables and state the remaining uncertainty. A claims-to-incident calibration requires a linked population with event observations, policy coverage and claim outcomes.
The practical role of exposure measurement is also to evaluate whether observed prevention outcomes are robust to portfolio change. Suppose a new control is deployed mostly to safer firms, and a fall in their observed rate follows. Selection, prior trends, changing detection, and time-varying exposures remain competing explanations. A carefully designed panel with baseline measures, timing, comparators and consistent ascertainment can strengthen analysis of control effectiveness. The present study supplies methodological implications, not a causal estimate for any security product.
For connected portfolios, incident generation and propagation need separate records. A provider's event frequency may be low while a large number of enterprises rely on its services. The portfolio's mean enterprise frequency does not describe the joint probability of simultaneous effects. Accumulation analysis requires dependency topology, coverage scope and conditional impact evidence in addition to frequency. Denominator relevance therefore extends beyond pricing individual risk to portfolio design and exposure monitoring.
A minimum comparable data release
A useful public statistical release would define its event taxonomy and version; identify whether attempts, confirmed compromise and impact are counted; disclose the eligible entity and coverage periods; report positive units and total event counts separately; publish denominator totals at the same granularity; state whether figures are occurrence-, detection- or report-dated; and document duplicate handling. It would disclose uncertainty and frame exclusions, including periods of incomplete observation.
Exposure series should be published alongside counts. At minimum an enterprise denominator needs entry and exit dates or an auditable approximation to enterprise-years, and asset-level studies need the relevant integrated active inventory. A fixed cross-tabulation of stable size and industry cells enables direct standardization. If security-sensitive granular data cannot be made public, aggregated cells with suitable suppression or privacy-preserving mechanisms can retain scientific usefulness. Privacy constraints should be disclosed rather than circumvented.
A companion metadata release should include survey or contributor changes, taxonomy changes, law and reporting-regime milestones, and data vintages. Versions permit retrospective reproducibility. Linkage information may be distributed in protected research environments where public disclosure would reveal sensitive incidents or infrastructure configurations. Independent replication should use authorized access and report aggregate results compatible with confidentiality obligations.
Such an approach makes interpretive disagreement more productive. Two researchers can compare results by asking whether they share event definitions, frame, exposure history and reporting assumptions. A disagreement about hazard can be isolated from a disagreement about what constitutes an event. Science advances when incompatible measurements are made explicit enough to reconcile or compare.
Future empirical research
A high-priority study would establish a longitudinal enterprise panel with stable technical and economic exposure records, linked occurrence and detection intervals, and periodic independent ascertainment audits. It would observe organizations that experience no incidents as well as cases, including entrants and exits. Repeated-event counts, prevalence and severity would be estimated separately by size, sector, technology dependence and event type. Public survey series could provide external descriptive calibration for eligible strata, without forcing identical definitions.
A second study would examine bridging between older and revised incident questionnaires. Randomly assigning alternative question forms to comparable respondent groups in the same reference period could estimate measurement effects under clear design assumptions. Re-administering a common core over successive waves would help preserve trend comparability. Such evidence would be especially useful when reported levels change at a documented questionnaire break.
A third study would investigate detection and reporting transitions. A consent-based sample of firms with independently audited monitoring records could be linked to internal response and administrative reports, subject to lawful confidentiality controls. Stratified estimates of detection and reporting probabilities could calibrate sensitivity bounds. This would make the gap between underlying incidents and visible reports a direct empirical question.
A fourth prospective study would establish common dependency identifiers to distinguish originating events from downstream impacts. With an exposure history of provider relationships, it could estimate conditional propagation frequencies and correlations. The hardest cases involve multiple overlapping providers, changing contracts and opaque subdependencies. These require careful data provenance rather than a single deterministic mapping.
A final research priority concerns comparability between insured and uninsured organizations. Secure matched data on enterprise incident events, coverage and claim decisions could estimate how policy scope, retentions and reporting behavior mediate claim record production. That research would support the allied study of insurance claim frequency while preserving independent enterprise incidence as the present monograph's object.
The proposed studies are research designs. They are not claimed as original named models or completed empirical investigations. Their significance follows from the identification gaps demonstrated by the current evidence.
Conclusion
Cyber incident frequency is a relationship between a defined event process and a defined quantity of exposure during a specified observation period. A count is a legitimate record of volume, while a prevalence measure, recurrence rate, asset-time incidence and policy-based ratio each represent a separate statistical estimand. Their interpretation depends on consistent classification, time attribution, eligibility and recording rules.
The available official evidence supports several scoped observations. Recent UK identified-attack prevalence fell between comparable early waves and was unchanged at the latest published rounding precision. Australia documents a break in its incident questionnaire. Canadian and European measures emphasize their selected business populations and consequences; their levels cannot be directly ranked against the broader UK measure. Administrative reports, the DBIR contributor network and ENISA's edition windows illustrate alternative observation mechanisms. US insurance market data demonstrate how growth in policy stocks can outpace growth in reported claims, without establishing a population hazard trend.
The worked mathematical analyses substantiate three conclusions. Exposure growth can reverse the implication of raw counts; shifts in population composition can reverse the implication of crude rates; and changing observation probabilities can reverse the implication of exposure-normalized recorded rates. These possibilities are distinct and may coexist. Neither additional precision in the numerator nor a more elaborate count distribution resolves missing information about exposure and ascertainment.
The science of cyber incidence requires measurement designs that preserve the correspondence between episodes, affected units, technical exposure, organizational time and the observation system. For insurance science, this discipline prevents public attack counts from being mechanically converted into claim frequency and ensures that cyber accumulation, contractual exposure and incident occurrence retain their separate empirical foundations.
A reproducible global rate remains contingent on an internationally comparable exposure and incident-record architecture that the accessible evidence does not yet provide. The scientific result of this investigation is a comparative, auditable interpretation of existing evidence and a program of empirical work directed at the remaining identification problems.
References
Australian Bureau of Statistics. (2023). Characteristics of Australian Business, 2021–22 financial year. https://www.abs.gov.au/statistics/industry/technology-and-innovation/characteristics-australian-business/2021-22
Department for Science, Innovation and Technology and Home Office. (2025). Cyber security breaches survey 2025. UK official statistics. https://www.gov.uk/government/statistics/cyber-security-breaches-survey-2025/cyber-security-breaches-survey-2025
Department for Science, Innovation and Technology and Home Office. (2026). Cyber security breaches survey 2025/2026: Technical report. https://www.gov.uk/government/statistics/cyber-security-breaches-survey-20252026/cyber-security-breaches-survey-20252026-technical-report
Department for Science, Innovation and Technology and Home Office. (2026). Cyber security breaches survey 2025/2026. UK official statistics. https://www.gov.uk/government/statistics/cyber-security-breaches-survey-20252026/cyber-security-breaches-survey-20252026
Edwards, B., Hofmeyr, S., & Forrest, S. (2016). Hype and heavy tails: A closer look at data breaches. Journal of Cybersecurity, 2(1), 3–14. doi:10.1093/cybsec/tyw003. https://academic.oup.com/cybersecurity/article/2/1/3/2736315
European Parliament and Council. (2022). Directive (EU) 2022/2555 (NIS 2 Directive). Official Journal L 333, 80–152. Article 23. https://eur-lex.europa.eu/eli/dir/2022/2555/oj/eng?uri=CELEX%3A32022L2555
European Parliament and Council. (2022). Regulation (EU) 2022/2554 on digital operational resilience for the financial sector. Official Journal L 333, 1–79. Article 19. https://eur-lex.europa.eu/eli/reg/2022/2554/oj?uri=CELEX%3A32022R2554
European Union Agency for Cybersecurity. (2025). ENISA Threat Landscape 2025. Publication record. https://www.enisa.europa.eu/publications/enisa-threat-landscape-2025
European Union Agency for Cybersecurity. (2026, September 22). ENISA Threat Landscape 2026. Publication record. https://www.enisa.europa.eu/publications/enisa-threat-landscape-2026
Eurostat. (2024, December 12). 93% of EU businesses apply ICT security measures. https://ec.europa.eu/eurostat/en/web/products-eurostat-news/w/ddn-20241212-1
Eurostat. (2026). Digitalisation in Europe — 2026 edition. ICT security section. https://ec.europa.eu/eurostat/web/interactive-publications/digitalisation-2026
Federal Bureau of Investigation. (2026). The Outreach — Spring 2026. Cryptocurrency and AI Scams Bilk Americans of Billions. https://www.fbi.gov/how-we-can-help-you/outreach/the-outreach/the-outreach-spring-2026
National Association of Insurance Commissioners. (2026). Report on the Cybersecurity Insurance Market. 2025 reporting year. Section 1, figures 3–5. https://content.naic.org/sites/default/files/2026-report-cybersecurity-insurance-market.pdf
Nelson, A., Rekhi, S., Souppaya, M., & Scarfone, K. (2025). Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile. NIST SP 800-61 Rev. 3. doi:10.6028/NIST.SP.800-61r3. https://csrc.nist.gov/pubs/sp/800/61/r3/final
Romanosky, S. (2016). Examining the costs and causes of cyber incidents. Journal of Cybersecurity, 2(2), 121–135. doi:10.1093/cybsec/tyw001. https://academic.oup.com/cybersecurity/article/2/2/121/2525524
Shevchenko, P. V., Jang, J., Malavasi, M., Peters, G. W., Sofronov, G., & Trück, S. (2023). The nature of losses from cyber-related events: risk categories and business sectors. Journal of Cybersecurity, 9(1), tyac016. doi:10.1093/cybsec/tyac016. https://academic.oup.com/cybersecurity/article/9/1/tyac016/7000422
Statistics Canada. (2024, October 21). Impact of cybercrime on Canadian businesses, 2023. The Daily. https://www150.statcan.gc.ca/n1/daily-quotidien/241021/dq241021a-eng.htm
Statistics Canada. (2024). Canadian Survey of Cyber Security and Cybercrime (CSCSC): 2023 reference period, survey 5244. Integrated Metadata Database. https://www23.statcan.gc.ca/imdb/p2SV.pl?Function=getSurvey&SDDS=5244
Statistics Canada. (2026, October 21). Impact of cybercrime on Canadian businesses, 2025. The Daily. https://www150.statcan.gc.ca/n1/daily-quotidien/261021/dq261021b-eng.htm
U.S. Securities and Exchange Commission. (2023). Cybersecurity Risk Management, Strategy, Governance, and Incident Disclosure. Small Entity Compliance Guide; final rule Release 33-11216. https://www.sec.gov/resources-small-businesses/small-business-compliance-guides/cybersecurity-risk-management-strategy-governance-incident-disclosure?trk=article-ssr-frontend-pulse_little-text-block
Verizon. (2026). 2026 Data Breach Investigations Report: report overview and data collection FAQ. https://www.verizon.com/business/resources/reports/dbir/