RESEARCH DOSSIER
Cyber Insurance External Attack Surface Scanning and Data Quality
Abstract. External attack surface scanning supplies information about externally observable digital resources, security conditions and technology dependencies used in cyber insurance underwriting. Its scientific value depends on the relation between a measured resource, an attributed insured organization, the time and scope of observation, and the contractual outcome that an insurer seeks to predict. This monograph investigates attack surface measures, measurement error and underwriting bias through source-critical analysis of network measurement research, original industry claims-correlation evidence, cybersecurity specifications, actuarial standards, an insurance policy specimen and United States and European governance sources. It distinguishes visibility, vulnerability, exploitation, incident, economic damage, insured claim and indemnity; formalizes the interpretative effects of missing observations, attribution uncertainty, prevalence, interval censoring and selection; and examines pricing, portfolio dependence, vendor accountability and proportionate remediation. Reproducible synthetic calculations show how false flags depend on base rates, how differential measurement error can produce disparate underwriting outcomes at equal latent prevalence, and how correlated losses alter portfolio variance without changing expected aggregate loss. The analysis supports context-specific, prospectively validated use of external measurements as one source of underwriting evidence. It establishes no vendor-specific performance estimate, newly named theory or original causal effect; public evidence does not permit independent reconstruction of private claims-correlation datasets.
Keywords: Cyber insurance; external attack surface; network measurement; security ratings; asset attribution; data quality; cyber underwriting; measurement error; missing data; calibration; selective labels; insurance claims; vulnerability severity; exploitation probability; portfolio accumulation; market conduct.
Introduction
Cyber insurance converts selected consequences of uncertain digital events into contractual and financial obligations. The relationship cannot be adequately studied by treating an observable vulnerability as a loss in itself. The relevant scientific problem is the construction of reliable evidence between the state of a technical system and the financial exposure of an insurance contract. This dossier investigates where external measurements contribute information, where they lose information, and how their uncertainty should be represented in underwriting, actuarial analysis and institutional governance.
The research follows three questions: how should external attack surface measures be defined and attributed; through which mechanisms do measurement errors change the interpretation of exposure; and how do such errors interact with selection, screening and the distribution of underwriting decisions? Its international scope is informed by public research and selected US and EU institutional sources. The relevant jurisdiction, document version, observed population and contractual boundary are stated for substantive claims.
The method combines critical comparison of technical and policy documents, analysis of original research designs, probability and exposure arithmetic, and institutional comparison. Synthetic examples use hypothetical inputs. The inquiry is an integrative monograph within Cyber Insurance, rather than an original empirical estimation exercise or a systematic review claiming exhaustive database coverage. The research cutoff is November 18, 2026.
Conceptual and Historical Foundations
The object of measurement
External attack surface scanning is the systematic observation of interfaces and indicators through which a digital system can be encountered from a specified external network vantage point. Its insurance value arises from the information it supplies about opportunities for compromise, operational practices and technology dependencies. The underwriting object is the uncertain contractual cost of a covered exposure. The two objects meet when a technical observation is attributable to an insured organization and credibly relates to a mechanism generating insured loss.
The external attack surface comprises publicly reachable or externally discoverable resources: domains, hosts, exposed services, public certificates, service configurations, and other signals describing an organization's internet presence. The observed surface is indexed by time, method, sensor location, permissions, and the population of resources searched. The actual exposure relevant to the organization additionally includes private networks, identity systems, employees, service providers, software dependencies, sensitive information, and business processes. An externally visible web gateway may be operationally critical, while hundreds of other public addresses may be largely irrelevant to the insured's business. A supplier can create material interruption exposure without appearing among addresses assigned to the policyholder.
An attack surface measurement describes an observation opportunity. An exposed service is a service observed to be reachable under defined conditions. A vulnerability is a weakness in an applicable implementation or configuration. Exploitability concerns whether the weakness can be used under relevant environmental and threat conditions. An incident is an occurrence affecting security or operations. Economic damage is a consequence borne by an organization or other affected party. An insured claim is a demand or reported loss presented under applicable coverage. Indemnity follows from the insuring agreement, other terms, facts, and adjustment. These units possess different denominators and cannot be collapsed into a single security score without sacrificing meaning.
Historical development and measurement economics
The 2013 ZMap research by Durumeric, Wustrow, and Halderman documented an experimental method for surveying public IPv4 space rapidly for a selected service. The ability to take repeated observations at scale enabled longitudinal network research. Its experimental throughput described a defined technical environment; it did not establish completeness across today's address families, services or organization boundaries. The development of outside-in underwriting followed a broader reduction in the cost of repeated external observations. It became feasible to collect observations about many proposed risks before a lengthy application or onsite assessment.
Current Censys research documentation describes distinct collections of public service observations, certificates and historical scan records. The separation matters scientifically because the existence of a certificate, an address and a functioning application are different facts. A certificate may remain visible after retirement; an address can change allocation; a gateway may represent multiple tenants. A measurement system must therefore preserve the relation between a raw observation and the attributed unit.
Lower information collection costs can reduce the cost of evaluating applicants, uncover discrepancies in questionnaires and permit risk engineering at a scale impractical through manual inspection alone. They also create incentives to standardize heterogeneous organizations into vendor ratings. Standardization can improve comparability only when the measurement process retains enough information about context. It can create false comparability when the same numerical value represents different coverage of the observed population, ages of the observations or mapping confidence.
The scientific distinction between prediction and assessment has practical significance. An internal control can be important even when it leaves no reliable external signature. Conversely, a visible anomaly may be predictive because it is associated with a wider pattern of organizational complexity, without itself being the initiating mechanism for the future incident. Underwriting should identify the role assigned to a feature: describing exposure, flagging a need for corroboration, estimating probability, determining eligibility, supporting loss prevention or informing accumulation analysis. Each use requires additional evidence appropriate to its consequences.
Security governance and the insurance boundary
Pascoe, Quinn, and Scarfone's (2024) Cybersecurity Framework 2.0 situates cybersecurity within risk governance and enterprise-wide outcomes. A perimeter observation is compatible with this framework as one information input. It does not demonstrate the implementation or effectiveness of governance, detection, response, recovery, or access management across an enterprise. Similarly, the official NIST testing guide treats technical tests as part of a structured assessment process. Underwriting data acquisition should be differentiated from authorized deeper security testing. This research performs no live scanning, accesses no private systems and provides no instructions for offensive exploitation.
The acquired signal and the resulting insurance transaction involve several distinct institutions. A measurement provider collects and processes observations. An attribution provider links resources with organizations. A security analyst interprets technical findings. A broker assembles application information. An insurer determines acceptance, terms and price. An actuary evaluates the relation to expected financial outcomes. A claims function evaluates what occurred and what a contract covers. The possibility that these functions are supplied by a single commercial group does not remove the need to distinguish their evidentiary roles.
The parent Science, Cyber Insurance, provides the general scientific context for coverage, exposure, loss measurement and portfolio dependence. The specific contribution of this dossier is to investigate the reliability and decision consequences of external measurement inputs, linking measurement science to underwriting and actuarial practice. It treats relevant contract terms as part of the target definition rather than treating the insurance contract as a generic promise against any manifestation of poor security.
Primary Documents and Scope of Attack Surface Measures
From observation to attributed exposure
An observation record should specify the resource identity, observation time, vantage category, protocol or finding type, measurement method version, result, expected coverage and status of verification. The linkage record should separately identify the proposed legal or economic entity, supporting evidence, effective interval, ownership or operational control relationship and ambiguity flag. The underwriting record should then link the attributed resource to the named insured or scheduled dependency, policy period, business function and relevant coverage. A many-to-many relationship is often the natural representation: one organization uses multiple resources, and a service provider or shared infrastructure can support multiple organizations.
Domain registration, certificates, network allocation records and observed service characteristics can suggest a relationship. They differ in their evidential force. A certificate name documents a represented name in a cryptographic credential; it does not alone establish legal ownership or present control of every resource. Address allocation can identify a network provider rather than the end customer. A corporate brand can be associated with several legal entities. Historical acquisitions and divestitures change the risk boundary independently of any change to a server. Consequently, a credible insured-level dataset stores both the observation and the reasoning by which attribution was made, along with dates and competing candidate matches.
An underwriting population requires a snapshot of an identified entity at a decision date. A later verified mapping may be useful in correcting a record, but it must not be silently inserted into the historical feature set used to evaluate prospective performance. In a study measuring actual decisions, the researcher needs the version available to the insurer when the decision was made. Later corrections should be recorded on a separate timeline. Reconstructing a perfect historical surface using current information can inadvertently introduce knowledge of events that occurred after the quote.
The expected denominator depends upon the question. Counts of observed open services may be suitable for describing a footprint. A fraction of successfully tested applicable assets exhibiting a finding describes conditional prevalence among measured assets. A rate of new findings per observed asset-time describes discovery or emergence, provided repeat observations are resolved into episodes. A count of unresolved observations at a renewal date describes a backlog under a particular matching rule. None of these is necessarily a count of unique exploitable paths.
Vocabulary for vulnerability evidence
The CVSS v4.0 specification (FIRST 2023) organizes vulnerability severity information into multiple metric groups. Its use can help communicate a weakness's technical characteristics. A severity score does not contain the monetary exposure of an insured. An organization-level assessment additionally requires applicability, reachability, effective controls, resource criticality and the circumstances of operation.
The EPSS model documentation (FIRST n.d.) describes a probability concerning exploitation activity for a vulnerability over a 30-day horizon. It distinguishes the task from technical severity and the historical evidence of known exploitation represented by KEV catalog inclusion. This predicted outcome is not the probability that a specific policyholder will be compromised in a given period. It is also not the probability that an insurer will receive a covered claim. The time horizon and unit of observation must remain attached to the value.
A feature extracted from vulnerability information should distinguish at least the candidate product identification, the confidence that the affected version is in use, the date of the vulnerability knowledge, the external path of reachability, the status of confirmation, the potential business consequence and the status of remediation. A patch may be backported without altering the visible version string. A network boundary may expose a benign presentation layer while the affected component is isolated. A vulnerability may be externally invisible because the vulnerable execution path depends on authentication or internal access. Conversely, configuration weaknesses may not map neatly to a vulnerability catalog identifier.
Attributes can also describe security practices independently of cataloged vulnerabilities. Recurrence of a misconfiguration, exposure duration, availability of a secure protocol option and malicious activity attributed to an organization may be informative. They require different time stamps and confirmation criteria. Observation of activity from an address can reflect infection, a shared service or erroneous attribution. A high rate of notifications can reflect better monitoring rather than worse operational performance.
Data representation and denominator design
The following table describes measurement units and interpretative limits. It supplies operational definitions for this research; no vendor is asserted to implement them in precisely this way.
| Measure | Denominator or unit | Principal qualification |
|---|---|---|
| Observed service count | Unique resource-service pairs at a snapshot | Duplicate front ends and dynamic allocation |
| Confirmed finding fraction | Affected tested applicable units / tested applicable units | Untested and inapplicable units are separate |
| Attributed footprint | Resources linked to an identified entity | Ownership, operation and dependency differ |
| Finding duration | Time bounded by repeated observations | Interval censoring and intermittent visibility |
| Scan availability | Completed eligible measurement opportunities / scheduled eligible opportunities | Unknown eligibility cannot be treated as success |
| Claim frequency | Covered claim counts / relevant earned exposure | Terms, reporting lag and exposure basis differ |
| Claim severity | Covered financial amount / specified claims | Deductibles, limits and development affect distribution |
The unit for a claim rate can be policy-years, insured-organization-years or another exposure appropriate to the contract. These denominators are not interchangeable where multiple policies cover one organization, where policies are layered, or where portfolio membership changes during a reporting period. A technical observation linked to an organization-year should not be counted repeatedly because that organization has several policies. Conversely, expected insurer payment may need to be evaluated policy by policy.
A scan report can be comprehensive relative to its own declared search scope while incomplete relative to the firm's business activities. Completeness should always be expressed against the intended measurement frame. In cases where the total relevant set is unknown, the analyst can report the known coverage and unresolved portions, without assigning an unsupported percentage of all business assets. The absence of a confirmed finding becomes negative evidence only within a specified tested and observable domain.
Contractual scope and evidence
The 2023 At-Bay policy specimen illustrates that different insuring agreements in a cyber contract can employ different event types and time triggers, and that direct and contingent business interruption are distinct insurance objects. Its aggregate limits and sublimits also structure the amount payable. The public wording is an example, not proof of any actual issued contract or market-wide uniformity. Insurers may use different definitions, waiting periods, conditions, exclusions and endorsements. A policyholder's observable vulnerability is therefore not a substitute for evidence of a covered network security event, interruption, liability or other specified loss.
Consider a hypothetical firm whose public interface is judged vulnerable and whose supplier later suffers an outage. The first fact concerns a posture signal, the second a possible dependent interruption. Establishing the economic loss requires business financial records, causation evidence and duration. Establishing entitlement requires the relevant insured identity, contract wording, declarations, policy period and loss category. The original scan may inform the history; it need not explain the outage. This example clarifies why a predictive feature may have actuarial value while lacking determinative force in the adjudication of a particular claim.
Mechanisms and Evidence for Measurement Error
Observation quality and uncertainty
The observed realization of a scan is conditional upon the existence of an applicable resource, access to the measurement opportunity, successful execution and the interpretation of the response. Missingness can occur at every stage. It may reflect resolution failure, address change, a temporary unavailable service, filtering, a new protocol, incomplete vendor coverage, or a gap in the process linking a resource to an entity. Such observations have different consequences. A negative result after successful applicable measurement is qualitatively different from a record with an unknown observation status.
Let T be a binary true condition for a specified unit and time under a reference assessment, and let O be a recorded binary finding for an eligible successfully observed unit. Sensitivity is P(O=1 | T=1), while specificity is P(O=0 | T=0). These describe measurement against an independent reference, not the probability of a covered loss. A complete evaluation also requires the share of eligible units successfully observed, the performance of attribution and the validity of the reference assessment.
If true-condition prevalence is p, sensitivity is s and specificity is c within a defined observed population, the expected positive fraction is s p + (1-c)(1-p). The positive predictive value for the condition is s p divided by that fraction when the denominator is nonzero. At p=0.05, s=0.80 and c=0.95, the expected flagged fraction is 0.0875; the positive predictive value is 0.04/0.0875, or approximately 45.71%. In a synthetic cohort of 10,000 observed units, the expected counts are 400 correctly flagged units and 475 false flags. There are also 100 missed true conditions. This illustration establishes the effect of prevalence on the evidential interpretation of a finding; it neither estimates actual vendor accuracy nor assumes that 5% is real prevalence.
When observation is incomplete, complete-case metrics describe the observed population. If the omitted units have systematically different characteristics, the observed rates generally cannot be transported directly. A known resource with an unsuccessful observation should retain an explicit unresolved status. A resource whose existence is unconfirmed requires an attribution or inventory status rather than a technical test status. Imputing both as secure masks uncertainty and changes the risk ranking by construction.
Time, cadence and duration
External observations expire in evidential value according to the rate at which the measured variable changes. A long-lived network allocation and a temporarily deployed application need not share a useful freshness threshold. A validated finding can become obsolete after remediation, a configuration change or disposal of the resource. A failure to repeat an observation can reflect the loss of visibility rather than remediation.
For a persistent finding first observed on day 10 and last observed on day 20, with valid negative observations on days 5 and 25, onset lies between days 5 and 10 and cessation between days 20 and 25 under an assumed single uninterrupted episode. Duration is bounded approximately by 10 to 20 days, not established as exactly 10 days. Multiple intermittent episodes would invalidate that simple bound. When there is no valid final negative test, duration is right-censored; when there is no previous negative observation, onset is left-censored. A model that interprets first detection as creation and last detection as resolution introduces systematic distortions linked to observation cadence.
A historical time series should differentiate newly detected assets from newly deployed assets, newly classified vulnerabilities from newly introduced weaknesses, and vendor extraction time from the time of actual observation. A software database update can retrospectively label a feature that was already present. If a training record includes future knowledge of that classification, performance estimates become optimistic for the stated prospective prediction task.
A stable entity-level feature can be constructed as the number of distinct observed relevant asset-days under a version-locked rule. In a longitudinal dataset, the numerator should not systematically increase simply because an insurer changes scanner vendors or purchases greater coverage. A vendor transition calls for a bridging sample measured independently by both systems under comparable conditions. Without such evidence, a change in score reflects an unidentified mixture of changing risk and changing measurement.
Signal validity and confounding
Liu and colleagues (2015), in Cloudy with a Chance of Breach, used 258 external features and historical incident reports from the period 2013–2014. The paper reports an operating point with sensitivity of 90% and a 10% false-positive rate for its classification task. That is evidence that observable characteristics can contain information about reported incidents. The operating point is specific to the study's design and its labels, and it cannot serve as an insurer's estimated annual claim probability.
In a hypothetical group of 10,000 firms with 1% annual incidence and a classifier operating at those illustrative rates, 90 of 100 positive firms and 990 of 9,900 negative firms would be flagged; precision would be 90/1,080 = 8.33%. The example transports the numerical operating point only as an arithmetic assumption to demonstrate the influence of prevalence. It does not transport the empirical result to a different time, sample, vendor, or insured outcome. Reporting overall accuracy without the event prevalence and classification threshold obscures this distinction.
Feature correlation is not a causal explanation. Firm size, industry, organizational complexity, technology choices, geographic presence and monitoring intensity can jointly affect external signals and incident reporting. A broad footprint can increase potential entry points and also reflect greater operational scale, redundancy, expertise and reporting capacity. A malware indicator can be an early warning or a consequence of an event already in progress. An adverse observation may prompt remediation before the policy begins; a later incident rate then reflects a changed risk state.
A simple statistical representation separates the observed scan variable X, latent security characteristics U, firm covariates Z, the applicant's decision to seek insurance A, insurer acceptance B, the occurrence of a covered event Y, and resulting payment L. Conditional relationships among these variables depend on the studied process. Fitting P(Y | X, Z, A=1, B=1) estimates a quantity for the selected book, given reliable outcome data; it does not identify the unconditional population risk, nor does it identify what would occur if an organization eliminated one feature through an intervention. Causal identification would require a defined intervention and assumptions that address competing explanations, selection, timing and exposure.
Evaluation of data fitness
The United States Actuarial Standards Board's (2016) ASOP No. 23 on data quality gives a professional framework for reviewing the relevance, reasonableness, sufficiency and limitations of actuarial data. Applied to scan inputs, the important question is whether a property of the data is adequate for the intended financial analysis. A fresh observation may still be attributed incorrectly; a correctly attributed historical record may be too old for a renewal assessment; a large dataset may omit a meaningful class of applicants.
A quality examination can be separated into three reference tasks. First, select a stratified sample of resource observations for independent confirmation under appropriate authorization. Second, evaluate entity linkage against documented organization boundaries and intervals, including acquisitions, vendors and ambiguous matches. Third, examine joins between the resulting entity histories and the insurance exposure dataset. Each stage should report counts, unresolved cases and reasons for exclusion. A single overall agreement rate cannot identify which stage contributes errors.
If two vendors agree because they draw on the same upstream data, their agreement need not provide independent confirmation. If they disagree because one observation is newer or the definitions of exposure differ, their disagreement need not mean that one vendor is generally inferior. Independent reference review should determine which proposition is disputed: existence, applicability, age, identity, severity or relationship to insured financial loss.
The audit record should retain the original evidence, the correction, its reason, the affected feature and the policies or decisions requiring reconsideration. A model that silently replaces an earlier value with the latest vendor score loses the ability to reconstruct what an underwriter knew and undermines retrospective testing.
Comparative Measurement of Underwriting Bias
Predictive evidence and market-specific limits
The 2024 Scanning the Horizon report describes Gallagher Re's linkage of firm data, externally derived indicators and material insurance claims. Its reported research sample includes more than 62,000 companies in 67 countries, more than 589 million IP addresses and over 1,000 claims meeting its materiality criterion. The report describes gradient-boosted modeling, benchmarks and comparisons with claims. This provides an unusually direct industry examination of the relation between external measurements and insured outcomes. It does not supply public claim-level data permitting the present author to reconstruct the fit, country-specific denominators, model uncertainty or effects on an independent insurer's portfolio. The involvement of a commercial vendor and broker is material to interpreting the scope and reproducibility of reported results.
The 2015 USENIX study by Liu and colleagues addresses public incident labels. Gallagher Re addresses insurance claims in a selected proprietary exposure universe. The two studies support the usefulness of investigating external signals for prediction, but their targets, time periods and data frames differ. They cannot be combined into a pooled numerical estimate of accuracy. Neither establishes that a single composite rating works in every market, product, firm size class, or underwriting period.
The 2026 Geneva Association cyber resilience report discusses limited visibility into internal architecture and the relationship between broker questionnaires, external indicators, demand for insurance and insurer screening. Its Box 9 reports broker-specific research covering 2019–2022. This evidence motivates separate study of application and acceptance decisions; it is treated here as an institutional synthesis of that research, not as a second independent replication or a verified reading of the underlying journal full text. The analysis also connects prevention and response services with the broader value of protection, making it important to evaluate how external findings are used rather than only whether they improve risk sorting.
Sampling and selection
Underwriting data emerge from a process with multiple filters. Potentially eligible organizations become known to brokers or insurers, some request quotes, some complete questionnaires, some obtain an external rating, some receive offers and some bind coverage. Claims are further conditioned on policy terms, events, recognition, reporting and loss development. A sample of paid claims supplies valuable observations for severity modeling; it cannot establish the claim frequency of organizations that never purchased coverage.
The selected sample can also obscure the value of a technical control. Suppose an insurer accepts only firms with few severe unresolved findings unless they document compensating controls. Among accepted firms with many findings, compensating controls will be unusually prevalent. A comparison between accepted firms with and without findings may then produce an attenuated or reversed relationship with later claims. This is a consequence of conditioning on acceptance and the insurer's behavior. It should not be treated as proof that the findings are harmless.
To study adverse selection, the researcher needs a defined market or solicitation population and evidence about application and take-up, not just risk scores among insureds. To study underwriting screening, the relevant denominator is submitted or eligible applications and the outcome is a specified decision, including declines, pending cases and conditional offers. To study insurer pricing, one needs offered prices as well as bound prices where possible, with coverage amount, retention, expected exposure and commission or other load differences. To study insurance losses, the denominator should be earned exposure with a specified claims valuation date. Combining these samples requires an explicit bridge between the stages.
An additional selection problem arises when identity matching is easier for firms with large public networks or well-established brands. A study restricted to successfully matched firms can overrepresent organizations whose online presence is easier to attribute. If, conditional on observed covariates, missing scores are still associated with unobserved risk characteristics, simple imputation or weighting may fail. Sensitivity analyses should vary unobserved risk among unmatched applicants and disclose the range of resulting portfolio-level estimates.
Underwriting bias and measurement disparity
Statistical bias in this context means systematic distortion of an estimated quantity relative to a specified target. Differential measurement error arises when confirmation, attribution or observation performance varies across relevant groups. Underwriting bias can also refer to unjustified or legally impermissible disparities in decisions. The scientific and legal meanings need to be evaluated separately. A group difference may reflect real differences in exposure or contract choice, measurement artifacts, inequitable procedures, or several mechanisms together.
For a synthetic example, assume two groups of 1,000 applicants have the same 10% prevalence of a particular true material condition. Both are measured at 80% sensitivity. Group A's specificity is 95%, whereas group B's specificity is 85% because infrastructure is more often shared and attributed ambiguously. Group A has 80 true flags and 45 false flags, while group B has 80 true flags and 135 false flags. Flag rates are 12.5% and 21.5%; positive predictive values are 64% and about 37.21%. An automatic decline rule based on the flags creates a nine-percentage-point difference despite equal underlying prevalence. A difference of that magnitude would merit an audit of the measurement process in this constructed example. It is neither evidence of a real protected-class disparity nor a conclusion regarding unlawful discrimination.
A score can be calibrated for the total population and miscalibrated within strata. A model might predict 4% annual claim risk in two groups, while observed developed claim frequencies are 2% and 6%. The pooled observed frequency is 4% if the groups have equal earned exposure. Portfolio-level calibration hides the subgroup discrepancy. A calibration assessment should compare consistent event definitions, comparable coverage and observation periods, reporting development, and uncertainty due to sample size. Differences should guide investigation rather than be automatically assigned a causal explanation.
The underlying group definitions should be linked to scientifically meaningful mechanisms: size, multinational structure, industry, hosting patterns, geographic measurement coverage, and product or contract type. The legal significance of protected attributes or proxies depends on applicable local law and the particular insurance transaction. Incorporating a protected characteristic into an audit for lawful bias assessment is a distinct question from using it to price or accept risks; access, legal basis and use should be controlled.
Out-of-time and external validation
A vendor score is decision-useful if it improves performance beyond information the insurer otherwise obtains and supports the intended action under relevant constraints. The relevant comparison is therefore not a model with scanning data versus a model that predicts the same average for every policy. An informative benchmark includes available firmographics, contract characteristics and verified application information. An incremental model adds versioned external indicators, under the same data cutoff, sample definitions and validation protocol.
The validation protocol should prevent leakage of related entities or repeated observations across training and test samples. Splitting the same organization-months randomly can produce favorable estimates due to memorization of stable identity attributes. A temporal split, in which features precede quotes and outcomes follow coverage, better reflects a prospective underwriting question. A second split by organization or corporate group tests generalization to new firms. A held-out provider or industry sample can test transportability where the purpose calls for it.
The ASOP No. 56 modeling standard (Actuarial Standards Board 2019) identifies the relevance of model purpose, appropriate data and overfitting to actuarial modeling. For insurance scanning models, discrimination metrics should be accompanied by calibration, frequency, severity, coverage and an assessment of decision cost. ROC area measures ranking across thresholds; a precision-recall curve better exposes the scarcity of positive outcomes when events are rare. The Brier score averages squared deviations between predicted probabilities and observed binary outcomes. None measures monetary value directly.
For a decision that combines the expected future claim cost with acceptance expenses, premium and capital consumption, a useful threshold depends on those costs and the proposed contract. The expense of false exclusion can be substantial: it includes foregone profitable business and the risk that an organization remains unprotected. The expense of false acceptance includes inadequately priced or accumulated exposure. A threshold derived on one book cannot be reused uncritically after changes in policy limits, risk appetite, underwriting controls or reinsurance structure.
Periodic validation is necessary because observable technology, attack behaviors, detection practices and the terms of cover change. The versioned workflow should define in advance what constitutes a material shift in feature availability, prediction calibration, subgroup error or score methodology. Analysis should separate a change in the risk distribution from a change in the observation system and from a change in underwriting acceptance. A model update should be evaluated with the same scrutiny as a new data source.
Financial Legal and Institutional Consequences
From a technical feature to expected claim cost
For one contract period, let N denote the number of covered claims, and let W_j denote the ultimate insurer cost for claim j after application of the relevant coverage terms. Aggregate expected insurer cost is E[sum of W_j for j from 1 to N | information available at underwriting]. The shorthand expected frequency multiplied by expected severity is appropriate when severity is defined consistently with the assumed count and dependence structure. For heterogeneous claims, one can instead sum expected costs over mutually specified claim types and exposures while avoiding double counting. A firm-level risk score only becomes financially interpretable through its empirically supported relationship with these quantities.
In a simplified single-loss example, suppose eligible ground-up financial loss X has a retention d and indemnity limit L above that retention, with no other applicable adjustments. The payment is min(max(X-d,0),L). For a synthetic ground-up loss of $300,000, retention of $50,000 and limit of $200,000, payment is $200,000 under that simplified assumption. The same technical event could produce different payments under other declarations, coverage triggers, sublimits or exclusions. Actual cyber coverage often includes multiple expense categories and complex aggregation; the equation is not a general legal interpretation.
Consider four synthetic exposures with equal assumed claim severity of $100,000 after policy terms. Two have annual claim probabilities 1% and 3% and two are uncertain because their data are missing. Assigning missing risks a probability of zero yields expected annual loss of $4,000. Allowing 1%–5% for each missing risk gives a range of $6,000–$14,000. Expected cost is additive regardless of dependence, but portfolio dispersion, capital requirements and tail exposure depend on joint outcomes. The calculation shows why treating unavailable data as favorable evidence is financially consequential. It is not a premium quote.
A reduced scan score may change the mean claim estimate, the uncertainty about that estimate, the underwriting action or none of them. A more defensible use of corroborated technical observations may be to trigger targeted risk engineering and obtain missing information. Automatically loading every uncertain exposure with a large price increment can introduce a systematic transfer of the insurer's measurement limitations onto applicants. The actuarial response to uncertainty should be documented and calibrated, with separate treatment of estimated risk, model uncertainty, expense and capital.
Aggregation and correlated loss
The resource graph can assist portfolio accumulation assessment when it identifies shared providers, common service types or technology dependencies. Entity-level observations alone are insufficient because multiple organizations can share an infrastructure provider that is invisible or ambiguously represented in the external resource dataset. Shared-resource attribution may be an error for one underwriting question and highly relevant for another. A provider-owned host should not be counted as independently controlled by every customer; dependency relationships should nonetheless remain available for scenario analysis.
Assume 100 insured firms face loss of $50,000 each if affected by one event. Under independent Bernoulli outcomes with characteristic probability 2%, expected aggregate loss is $100,000 and variance is 100 × 0.02 × 0.98 × 50,000 squared, or $4.9 billion in squared-dollar units. Under a hypothetical common shock that affects all 100 firms together with probability 2%, the same expected loss of $100,000 accompanies variance 0.02 × 0.98 × 5,000,000 squared, or $490 billion in squared-dollar units. The standard deviations are $70,000 and $700,000, respectively. These extreme cases illustrate how dependence changes aggregate distributions without changing their means.
Insured aggregation must account for financial terms, attachment and exhaustion of layers, reinsurance recoveries, the distinction between direct and dependent interruption, and potential shared exclusions. Technology similarity is a scenario input rather than a sufficient estimate of common-shock probability. A portfolio might share a provider but contain diverse recovery arrangements and coverage. A scenario can be useful even with uncertain probabilities provided it identifies the modeled event, insured dependencies, loss assumptions, uncertainty and objectives.
EIOPA's cyber insurance overview distinguishes risks to the insurer's own operation from risks accepted through cyber coverage. That distinction also applies to a vendor supplying scan data to an insurer. A vendor compromise or outage can disrupt the insurer's operations or impair decision quality, while an event at a technology provider used by insureds can generate contractual liabilities. An overlapping service provider can create both categories of risk. Governance should record the relevant mechanism and avoid merging them into one unexplained risk amount.
Regulatory and professional context
The 2026 NAIC cybersecurity insurance report covers the 2025 reporting year using United States insurer reporting and specified surplus-lines information; its opening methodology differentiates primary, excess and endorsement reporting. These are important dimensions of insurance exposure and market experience. Aggregate market statistics are not a linked public sample of each insured's external scan, contract and developed claim. They can contextualize a book's segmentation while leaving the individual-level relationship unidentified. A comparison with other national markets requires harmonization of reporting definitions before drawing conclusions.
In the United States, the applicability of insurance law depends on the jurisdiction and product. The New York Department of Financial Services' (2024) Circular Letter No. 7 sets supervisory expectations for the use of AI systems and external consumer data in underwriting and pricing, including oversight of vendors and correction of unreliable information. Its treatment of data accuracy, insurer responsibility and adverse-decision transparency offers a specific governance example. Scope must be assessed using the circular letter's definitions and relevant New York insurance statutes; a commercial network observation about a corporate entity cannot automatically be treated as consumer data in all cases, nor does the circular letter impose the same requirements in other jurisdictions.
In the European Union, the General Data Protection Regulation (European Parliament and Council 2016) addresses accuracy and purpose limitation for personal-data processing, alongside lawful processing requirements. A network observation is not necessarily personal data; some observations or combinations of records can relate to identifiable natural persons. Applicability requires evaluating the type of data, roles of controller and processor, context and territorial scope. Corporate security information may also be sensitive as a commercial matter independently of personal-data rules. Neither a public-facing interface nor a vendor's access to observations establishes permission to collect, store or redistribute all associated information.
Professional actuarial guidance and cybersecurity standards occupy different institutional roles from law and individual contract provisions. They inform reasonable measurement, modeling and documentation, within their stated scopes. Regulatory decisions depend on the relevant supervisory and statutory frameworks. The international comparison in this research is therefore a comparison of several documented governance layers rather than an exhaustive account of every jurisdiction's law or market practice. It establishes no worldwide legal duty to purchase or use a particular scanning product.
Supplier accountability and operational safeguards
A data supply agreement should specify the intended use, observation coverage, update and correction processes, access to supporting evidence, version history, acceptable use, incident notification and procedures for service termination or transition. Insurers should establish which data can be retained for reconstructing decisions and which must be deleted or access-restricted. The terms need to be consistent with applicable law and the insurer's own obligations.
A mechanism for disputing attribution is valuable for both accuracy and market conduct. An applicant should have a reasonable opportunity, under an appropriate workflow, to explain that a recorded resource belongs to another entity, has been retired, is hosted by a provider, or has undergone remediation. An insurer should retain the original record, corroborating evidence, revision and decision rationale. Appropriate disclosure should avoid publicly distributing sensitive details of a firm's unresolved exposures or information irrelevant to the decision.
Loss prevention and underwriting can produce different incentive effects. Revealing credible findings to an authorized insured can support remediation. Penalizing every reported weakness regardless of confirmation, context or corrective action can discourage candor and create incentives to optimize appearances. A measured risk improvement should concern a documented change in meaningful exposure or controls, rather than simply a change in public visibility. Score-driven incentives need evaluation as part of the underwriting process because the process itself can alter the data-generating mechanism.
Integrated Scientific Synthesis and Research Agenda
A research program for decision-useful measurement
The evidence supports three related conclusions. External security observations can contain predictive information about selected cyber outcomes. Their significance depends on measurement, attribution, timing, and the insurance outcome against which they are tested. A reliable decision requires evidence at both the technical and contractual levels, linked by a fit-for-purpose statistical design. These conclusions synthesize existing network measurement, insurance, professional and governance sources. They constitute an analytical integration rather than a newly named theory.
A prospective cohort study is a suitable next step. Researchers would identify the applicant and covered-population frames, define underwriting decision dates and coverage periods, specify policy-level risk and claims outcomes, collect original feature vintages, and record all exclusion reasons. A separate validation sample should include independent confirmation of technical findings and legal-entity linkage. The data record should distinguish successful observation, failed observation and inapplicability; disputed mappings remain explicit. An anonymized research identifier can link multiple reports without disclosing detailed sensitive asset information in the publication.
For event prediction, the outcome might be a first reported covered claim during a contract period. The observation window and valuation date should be fixed. If the objective is ultimate expected cost, the study should adjust for claim development and include cases with no reported payment. A portfolio can experience changes in severity distribution without changes in claim frequency. Separate outcomes therefore support a more informative analysis than a single composite incident measure. A study of loss prevention should instead define the intervention, adherence, spillovers, comparable control groups and potential alternative explanations.
The planned baseline can use firm size, sector, location as appropriate, verified questionnaire variables and contract structure. A second specification can add individual external observation features. A third can explore the role of missingness and attribution confidence. All must be assessed under the same data partitions and outcome definitions. Researchers should report absolute and incremental calibration, ranking, financial decision metrics, error patterns and uncertainty. They should predefine the treatment of duplicated organizations, interrelated exposures, zero-payment claims, open claims and changes to policy terms.
An independent assessment of vendor agreement should compare matched low-ambiguity resources as well as broad applicant coverage. Agreement in a curated sample may differ from agreement in the entire underwriting population. Adjudication of sampled discrepancies can identify whether the disagreement reflects vintage, coverage, technical classification, attribution or scoring method. A finding that one provider performs better for a narrow specific feature does not establish overall superiority. Combining providers can increase observed coverage but can also increase correlated errors and processing costs.
Reproducibility and evidence sufficiency
A reproducible analysis requires a data dictionary, stable version identifiers, documented joins, sampling dates, inclusion criteria, exposure definitions and sensitivity analysis. Where raw data are commercially sensitive or unsafe to publish, the public research package can still contain aggregated validation results, reproducible synthetic examples, a description of access conditions and a record of analytic choices. Disclosure restrictions should not be hidden by reporting spurious precision.
For each empirical number cited in this monograph, the reference identifies the source population and target. The original academic work reports historical incident classification; the industry report describes a particular matched exposure dataset and material claims. Neither has been independently replicated in this study. The official insurer market report describes reporting categories and serves as market context rather than as evidence of score accuracy. The technical specifications establish terminology and numerical meaning rather than insurer outcome calibration. Legal and professional instruments are differentiated by jurisdiction, institutional authority and scope.
The literature retrieval concentrated on public original documents, official publication records and a relevant institutional synthesis available through November 18, 2026. Some paywalled or inaccessible documents and primary incident datasets could not be inspected directly. The underlying individual-level broker study discussed in the Geneva Association report was not independently accessible, so the monograph attributes the discussion to that report and makes no coefficient-level claim. The CISA catalog webpage was inaccessible directly; its general distinction from EPSS is referenced through FIRST's official explanation without claiming a live count. The ENISA 2026 publication page could not be substantively retrieved in this session, and no distinct numerical result from it is asserted. PDF visual checks succeeded for the selected sample policy page and ZMap opening page; source text, rather than visual inspection, was available for the cited Gallagher and NAIC excerpts. These limits define the evidence population used here.
No private claim-level or external scan dataset was accessed, no live network was evaluated, and no commercial scoring methodology was independently audited. The numerical examples are deterministic illustrations, not estimates of current market behavior. Their inputs are deliberately transparent and their outputs recomputable. A scientific conclusion about the amount of incremental explanatory power provided by a named supplier remains conditional on validated data for the relevant insurer and product.
Implications for insurer governance
A rational underwriting process can assign graduated evidential weight to observations. Verified, timely and attributable conditions can enter a validated model or support proportionate risk engineering. Plausible but unresolved observations can trigger inquiry. Unobservable areas can produce a request for alternative evidence or an explicit uncertainty allowance. An unsupported negative observation should not silently become proof of effective security. The treatment should be justified by the intended decision and its financial and institutional consequences.
For an insurer, the acquisition of scanning data is a continuing obligation of integration and evaluation, not merely a procurement decision. External observations can assist portfolio screening and inform discussions with policyholders, but their value is mediated by the processes that turn them into attributed features, actuarial estimates, risk decisions and corrections. A record of those processes allows a later reviewer to establish whether the information available at the time was suitable and whether new evidence warrants a correction.
This analysis also identifies the importance of maintaining distinct controls for claim adjudication. A historical scan may establish evidence of a reported external condition, but a subsequent claim requires a specific factual and contractual assessment. A low historical score is not proof of the cause of an event. A high score is not evidence of entitlement absent the facts and terms of coverage. In a dispute, original observations, times, linkage, policy wording and causation are each subject to their own evidentiary requirements.
The principal research priority is access to comparable prospective cohorts with transparent exposure histories, relevant contract terms, and developed financial outcomes. Without them, published rankings and retrospective fits can demonstrate plausibility without resolving transportability. Transparent measurement uncertainty and outcome-specific validation provide more scientific value than a superficially precise, generalized claim about a vendor's universal ability to predict cyber losses.
Conclusion
External attack surface measurements are observations of a defined, changing technical environment. Their strongest scientific contribution to cyber insurance is the provision of evidence about conditions that may affect covered losses, subject to attribution, timing, applicability, validation and contractual interpretation. Their limitations are similarly specific: externally inaccessible controls, dynamic resources, shared infrastructure, uncertain matches, incomplete labels, selection into the insured book and changing exposure terms.
Original network measurement research establishes the feasibility of frequent large-scale observation and the predictive relevance of certain historical external features. Original industry work indicates that these signals can also be useful in a matched insurance exposure setting. The evidence cannot establish universal calibration, vendor interchangeability or the financial effect of a particular remediation intervention. An insurer's prospective validation must address the relevant population, coverage, exposure, measurement vintages and loss development.
Measurement quality affects both estimation and decision fairness. The synthetic examples show that sensitivity and specificity require interpretable denominators, false flags depend on prevalence, differential specificity can produce unequal outcomes at equal true-condition prevalence and dependence can magnify aggregate variability. Accordingly, a reliable decision retains uncertainty and distinguishes unobserved conditions from verified favorable conditions.
The scientific contribution of this dossier is a structured cross-disciplinary analysis connecting measurement validity, underwriting selection, contractual loss and portfolio dependence. It explains why the value of external data is a property of an entire evidence-to-decision process, including correction and governance. Its conclusions remain bounded by the accessible sources and transparent hypothetical calculations. No new scientific doctrine, claim of independently estimated empirical effects, or universal legal rule is advanced.
References
Actuarial Standards Board. 2016. Data Quality. Actuarial Standard of Practice No. 23. December 2016. United States. Publisher or original document
Actuarial Standards Board. 2019. Modeling. Actuarial Standard of Practice No. 56. December 2019. United States. Publisher or original document
At-Bay. 2023. Cyber Insurance Policy. Form AB-CYB-001.2, edition 08/2023, pp. 2–4. Publisher or original document
Censys. n.d. Research Access to Censys Data. Product documentation. Accessed November 18, 2026. Publisher or original document
Durumeric, Zakir, Eric Wustrow, and J. Alex Halderman. 2013. ZMap: Fast Internet-Wide Scanning and Its Security Applications. Proceedings of the 22nd USENIX Security Symposium, 605–620. USENIX Association. Publisher or original document
European Insurance and Occupational Pensions Authority (EIOPA). n.d. Cyber Insurance. Institutional overview. Accessed November 18, 2026. Publisher or original document
European Parliament and Council of the European Union. 2016. Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). Official Journal L 119, 1–88. Publisher or original document
Forum of Incident Response and Security Teams (FIRST). 2023. Common Vulnerability Scoring System Version 4.0: Specification Document. Accessed November 18, 2026. Publisher or original document
Forum of Incident Response and Security Teams (FIRST). n.d. Exploit Prediction Scoring System: Frequently Asked Questions. Accessed November 18, 2026. Publisher or original document
Gallagher Re. 2024. Scanning the Horizon: How Broadening Our Use of Cybersecurity Data Can Help Insurers. Claims-correlation research report prepared using Bitsight data. October 2024. Publisher or original document
Liu, Yang, Armin Sarabi, Jing Zhang, Parinaz Naghizadeh, Manish Karir, Michael Bailey, and Mingyan Liu. 2015. Cloudy with a Chance of Breach: Forecasting Cyber Security Incidents. Proceedings of the 24th USENIX Security Symposium, 1009–1024. USENIX Association. Publisher or original document
National Association of Insurance Commissioners (NAIC). 2026. Report on the Cybersecurity Insurance Market: 2025 Reporting Year. October 2026. Publisher or original document
New York State Department of Financial Services. 2024. Insurance Circular Letter No. 7: Use of Artificial Intelligence Systems and External Consumer Data and Information Sources in Insurance Underwriting and Pricing. July 11, 2024. Publisher or original document
Pascoe, Cherilyn, Stephen Quinn, and Karen Scarfone. 2024. The NIST Cybersecurity Framework (CSF) 2.0. NIST CSWP 29. National Institute of Standards and Technology. DOI: 10.6028/NIST.CSWP.29. Publisher or original document
Scarfone, Karen, Murugiah Souppaya, Amanda Cody, and Angela Orebaugh. 2008. Technical Guide to Information Security Testing and Assessment. NIST SP 800-115. National Institute of Standards and Technology. DOI: 10.6028/NIST.SP.800-115. Publisher or original document
The Geneva Association. 2026. Strengthening Cyber Resilience Through Insurance. Research report; section 4.2.1 and Box 9. Accessed November 18, 2026. Publisher or original document