Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Generative AI Phishing and Cyber Insurance Social Engineering Losses

Abstract. This Research Dossier investigates generative AI-assisted phishing and social engineering as sources of cyber insurance loss. It integrates experimental studies of persuasive communications, official incident and complaint evidence, insurer-reported claims categories, selected commercial crime coverage decisions, technical identity standards and payment reimbursement rules. The analysis separates attack capability, exposure, deception, transaction execution, recovery and contractual coverage. It demonstrates why click rates, complaint losses and paid insurance claims have different denominators, and why the attribution of AI use cannot substitute for legal causation. Conditional-probability, indemnity and dependence illustrations examine the consequences of alternative assumptions without claiming empirical calibration. The work develops an original analytical synthesis of the established research and identifies data requirements for defensible underwriting, claims classification and portfolio analysis. International conclusions are bounded by the observed populations and the specific legal documents examined.

Keywords: generative AI; phishing; social engineering; cyber insurance; business email compromise; funds transfer fraud; cyber claims; insurance contract causation; risk engineering; payment fraud; loss attribution; actuarial analysis

Introduction

Social engineering converts an organizational decision into a source of loss. Communication, identity, delegated authority, financial authorization and recovery are jointly involved in the resulting exposure. Generative AI changes the production and adaptation of deceptive communications, while insurance responds to contractually defined events and consequences. The subject of this dossier is the relationship between those changes and insured fraud losses, specifically in the context of enterprise phishing and impersonation. Broader AI-enabled cyber threats, autonomous malware and general coverage of AI liability belong to adjacent research.

The investigation evaluates three questions: the evidence establishing AI involvement; the influence of scalable persuasion on exposure and control failure; and the classification of incidents into insurance claims with identifiable monetary consequences. The principal conclusion is that technology attribution, the causal route to loss and the contractual grant of coverage must be recorded separately. Aggregation across these dimensions requires a common unit of observation. The scope is international in empirical perspective, with specifically identified US insurance decisions, UK payment reimbursement and Hong Kong incident documentation; it does not purport to state the insurance law of every jurisdiction.

The research is a structured documentary synthesis as of November 18, 2026. Sources were selected for direct relevance to the defined object: complaint statistics, insurer-reported claims summaries, intelligence assessments, technical standards, original reported court opinions and empirical research. Each source's population and endpoint are preserved. The design uses comparative legal interpretation, conditional probability, loss-layer arithmetic and explicit sensitivity analysis. It is not a systematic review with a registered search protocol, and no private insurer microdata or paid-claim file is available. Illustrations are hypothetical and have no implied calibration to the sources. No new named theory is advanced.

Conceptual and Historical Foundations

The insured event and its precursor

Phishing is a deceptive communication intended to induce an action in reliance on an apparent legitimate identity or purpose. Spear phishing is the targeted form. Social engineering is broader: it includes other manipulations of human decision-making across digital and nondigital channels. Business email compromise (BEC) describes fraud conducted through abuse of organizational email identities or processes; funds transfer fraud (FTF) describes a monetary outcome or contractual category, depending on the source. Neither term is a reliable synonym for generative AI use.

An email may precede a credential disclosure, a compromised mailbox, a payment instruction, a privacy breach or an interruption of operations. The ultimate loss may arise from an unauthorized account takeover or a payment issued by an authorized employee who was deceived. The initial communication, access event, financial event and insurance claim are different observation units. A single communication can yield no loss; a single compromise can yield multiple demands; several transfers can become one insurance claim. Conversely, one incident can implicate more than one policy or insured party.

This distinction predates generative AI. Ho et al.'s (2019) study of lateral phishing examined enterprise email data from 92 organizations containing 113 million employee-sent messages, establishing the importance of compromised internal accounts as carriers of preexisting trust. An authentic account can convey a fraudulent instruction. The resulting security problem is therefore more extensive than detecting whether a text is artificially generated. Identity authenticity and authority for a specific transaction must each be established.

Organizational authority and contextual credibility

Employees make decisions within a division of labor. Accounts payable processes new instructions; procurement maintains supplier records; treasury manages approval and settlement; information technology administers credentials. Errors can be produced when a communication appears consistent with one role's legitimate obligations while bypassing controls owned by another. A familiar writing style is a weak basis for establishing authority because a legitimate sender may change style and an impersonator may reproduce it.

Steves, Greene, and Theofanos (2020) analyze audience-specific message difficulty in their Phish Scale study. Their finding that alignment with work context changes difficulty supports comparing simulations by scenario and recipient role. A declining unadjusted click rate can reflect easier tests rather than stronger controls. Training is a useful observational setting when interpreted with context, but its endpoint is distinct from monetary fraud.

The operative boundary is the insured interest. This may be the insured's own funds, its legal liability to another party, expenses incurred in responding to compromise, or a specified business interruption loss. Ownership, contractual responsibility and insurance wording determine which measures are relevant. A customer deceived into paying a criminal rather than the insured may create a collection dispute, reputational cost or liability allegation even when no insured-owned money was taken. A loss database must distinguish that case from theft of funds held by the insured.

Historical continuity and change

Generative systems alter the cost, timing and variety of deceptive content. The underlying authority problem existed in earlier fraud: false instructions, impersonated counterparties and misleading communications. A useful insurance baseline is thus conventional social engineering exposure, assessed before estimating an incremental effect from AI. The emergence of a new capability changes both prevention requirements and data uncertainty without automatically creating a novel insuring agreement. An insurer that treats every contemporary BEC event as AI-driven will confuse chronology with evidence; an insurer that ignores AI when supported by incident evidence loses an explanatory variable of potential underwriting relevance.

AI-Enabled Phishing

Empirical evidence and its limits

Heiding et al. (2024), in their human-subject phishing preprint, allocated 101 participants among four message groups. Rounded click proportions were 12% for the control, 54% for human experts, 54% for fully automated messages and 56% for AI-assisted human intervention. The variants differed in personalization depth, so this experiment compares composite approaches rather than isolating the independent causal effect of text generation. It offers evidence that automation can produce persuasive communications in a specific experimental setting. Small group sizes, the recruitment population and the surrogate endpoint limit its transfer to corporate payment authorization. The difference between 54% and 56% should not be interpreted as a reliable ranking.

The conference abstract by Czybik et al. (2026), A Large-Scale Study of Personalized Phishing using Large Language Models, reports an experiment with 7,700 participants and a near-tripling of clicks for personalized LLM communications compared with generic approaches. This larger study supports taking contextual personalization seriously. Only the conference abstract is relied upon here; detailed design, subgroup effects and uncertainty intervals are not inferred from it. Both studies measure interaction with a message, not the conditional likelihood of an insured transfer. Their joint relevance lies in the early stages of a loss process rather than direct calibration of insured severity.

The United Kingdom NCSC's January 2024 near-term AI threat assessment anticipated difficulty differentiating legitimate requests from deceptive communications. Its subsequent assessment through 2027 projects expanded effectiveness of intrusion operations and uneven capacity for defense. These are probabilistic intelligence judgments. They strengthen the case for forward-looking scenarios, but they supply no insured-frequency multiplier.

Text, audio, video and the evidentiary boundary

An incident can combine several communication media and sources of apparent authority. The June 26, 2024 Hong Kong legislative reply documents a January case involving an impersonating email and a fabricated prerecorded video conference, followed by payment authorization and approximately HK$200 million in loss. At the reporting date the police investigation remained open. The official account specifically describes the conference as prerecorded, so the event does not establish an interactive real-time deepfake or autonomous attacker. Nor does it establish insurance coverage, insurer participation or the amount ultimately recovered.

For claims investigation, use separate labels for suspected generation, corroborated use and unknown technology. Direct evidence could include preserved original artifacts with a reliable chain of custody, trustworthy process logs, findings documented by qualified forensic examiners or authenticated admissions. A detector score alone is insufficient: it depends on its evaluation population, version, channel and error properties. A polished message, unusual accent, or inconsistent video frame may justify further inquiry; these characteristics are not a conclusive causal classification.

Even proven synthesis has several possible functions. It may have generated the first contact, supported an impersonation during payment approval, assisted translation or contributed to another stage without affecting the eventual decision. For insurance research, record the stage and the evidence for its connection to the loss. A proven generated image included in an otherwise unrelated exchange does not establish that AI caused the transfer. Conversely, the lack of a recoverable generated artifact cannot establish that the actor used conventional methods.

Human and institutional outcomes

A technically authentic digital session can be associated with a fraudulent transaction. Authentication establishes that a claimant or device satisfies a protocol. Transaction authorization concerns an instruction to move a specified amount to a particular beneficiary for an identified business purpose. Proper separation of these objectives limits loss where authentication succeeds but a trusted person is manipulated. This separation is central to interpreting technical controls in insurance questionnaires. A binary control field, such as the presence of multi-factor authentication, describes very little unless its deployment coverage and the process being protected are identifiable.

The European Union Agency for Cybersecurity's 2026 threat landscape summary uses observed incidents and events in 2025 and identifies social engineering and dependency exposure as persistent concerns. Its observations are relevant to changing threat pathways across organizations. Its source mix and reporting boundary preclude interpreting the recorded events as claims per insured enterprise-year.

Attack Scalability and Risk Engineering

Conditional loss pathways

For analytical purposes, let a loss opportunity be a distinct payment or access decision exposed to a deceptive instruction. If A denotes exposure to that instruction, B successful deception, C execution of a consequential action, D failure to prevent or reverse the loss, and I satisfaction of the policy's insuring conditions, a joint probability can be decomposed using conditional probabilities:

P(A, B, C, D, I) = P(A) P(B | A) P(C | A, B) P(D | A, B, C) P(I | A, B, C, D).

This identity makes no independence assumption. The insured event depends on the complete path; a change in any conditional stage may increase or decrease risk. A message click is an imperfect measure even of B because an employee may inspect a message without believing it. Economic intervention at C or D can substantially reduce realized loss even where persuasive messages proliferate. Coverage I is a separate legal and financial condition, rather than a security outcome.

Scale can affect the number of exposures, the dependence among events and the distribution of reachable transaction amounts. If multiple messages concern one decision, counting each as an independent exposure inflates predicted event counts. A campaign may affect multiple clients of a supplier; common dependence can then persist even where each company has a distinct risk policy. In the reverse direction, improved detection and recovery may offset parts of the change in external pressure. No single fixed multiplier follows from generative capability.

A hypothetical population of 1,000 distinct annual exposed decisions illustrates the distinction. Suppose 4% of decisions yield consequential fraudulent actions and 25% of those actions remain economically unrecovered; suppose 60% of unrecovered events qualify under a given contract. The expectation is 6 covered events. In a second scenario exposure grows to 1,500 decisions, the consequential-action rate remains 4%, the unrecovered share falls to 10% and coverage remains 60%. Expected covered events decline to 3.6. The example holds other features constant by construction: it is a demonstration of conditional arithmetic, not an estimated AI effect. Event counts need not be integers in expectation.

Defense as institutional design

NIST's (2025) phishing-resistance requirements distinguish authentication protocols that protect secrets from impostor verifiers from manual-output mechanisms that can be relayed. This supports underwriting questions about whether sensitive accounts use phishing-resistant authentication, including effective enrollment, recovery and exceptions. It does not resolve the risk of an authorized employee issuing a deceit-induced payment.

NIST's CSF 2.0 guidance addresses governance alongside other security outcomes. Applied to organizational fraud prevention, governance assigns ownership of controls, identifies critical financial workflows, monitors exceptions and makes response responsibilities clear. A relevant defensive program can incorporate independently maintained supplier records, authenticated verification through established contacts, separation of payment initiation and approval, review of material beneficiary changes, transaction anomaly monitoring and a rapid contact path to payment institutions. These are risk engineering suggestions, not a universal legal condition or proof of effectiveness for a particular contract.

Segregating two approvers is valuable only insofar as their decisions are meaningfully independent. If both rely on the same unreliable communication, the nominal redundancy can fail. A useful audit therefore examines the information available to each approver and whether an alternate verification method is independent of the disputed instruction. The control should be tested against authorized simulations without collecting unnecessary employee information.

Loss prevention has a recovery dimension. A report can result in a stopped transfer, recalled funds, recovery after settlement or a partial return. An organization's response readiness should support prompt communication with financial institutions, forensic preservation and compliance with applicable policy notice provisions. The recovery rate must be measured with a fixed maturity window; an insurer's open claims and recently detected incidents have had less time to mature than older records.

Portfolio dependence

Consider two equal-sized insured exposures with hypothetical claim indicator probability p = 0.01. Under independence, the probability both incur a claim is 0.0001. Under a common shock with perfectly coincident indicators, it is 0.01. Expected count remains 0.02 in both cases. The variance of the count changes from 0.0198 to 0.0396. The example demonstrates that marginal frequency alone cannot represent portfolio concentration. Real portfolios lie between many possible dependence structures, and the common-shock case is an illustrative extreme.

Dependency mapping should identify common payment platforms, identity providers, shared service organizations and significant vendors using lawful, proportionate data. Not every common provider produces a correlated insured event; the dependency must connect to the relevant loss pathway. Campaign-level exposure, affected policyholders, loss components and contractual occurrence rules require different identifiers. Reinsurance aggregation must be tested under the actual wording rather than assuming all losses associated with similar synthetic material form one occurrence.

Incentives and operational proportionality

Risk classification may affect underwriting terms, access to insurance and incentives to invest in controls. A control warranty can shift economic consequences sharply if it operates as a prerequisite to payment. An observable and specific question is more useful than a vague representation of AI readiness: for instance, scope of independent payment verification, exceptions and historical recovery experience. Proportionality is important because smaller firms may have different staffing and transaction workflows. Equivalent outcomes can be supported by different organizational designs. Rates and coverage decisions should reflect relevant measured risks without treating technology adoption or workforce characteristics as automatic proxies for vulnerability.

Claim Classification and Insurance Consequences

Observed complaint losses versus insured claims

The FBI's 2025 IC3 Annual Report (published 2026, pp. 26 and 39) gives BEC complaint losses of $2,946,830,270 in 2023, $2,770,151,146 in 2024 and $3,046,598,558 in 2025. The 2025 increase over 2024 is approximately 9.98%. The same publication records 22,364 complaints containing AI-related information and $893,346,472 in adjusted losses across that wider descriptor, with more than $30 million in reported BEC losses involving AI. These are complaint-based figures, and the wider AI descriptor spans crime categories. The 2025 AI-related BEC figure is not a complete, validated market prevalence estimate. Adding it to total BEC would risk overlap. A rise in reported losses can reflect changing attack activity, severity, identification, reporting or classification.

Coalition's 2026 claims report summary describes its own portfolio of more than 100,000 policyholders during 2025. BEC and FTF together accounted for 58% of all claims, and 71% of FTF claims were attributed to social engineering; 52% of FTF claims arose from BEC. These shares have distinct denominators, and the overlap between the latter two proportions is not specified. They cannot be added or treated as nationwide proportions. The publicly accessible summary is used here rather than an inaccessible full report. The evidence establishes the importance of the loss categories for that carrier, not an AI-specific insurance payment total.

For comparability, record exposure in matched policy-years, report dates, event dates, claim maturity, gross economic loss, claimed amount, final indemnity, expenses and recoveries. Claims reported and claims paid are not interchangeable. For international observations, preserve original currency, any chosen conversion rate and conversion date; do not combine nominal amounts from different jurisdictions without a defined purpose and adjustment. The Hong Kong case is retained in HK dollars.

Coverage architecture and causation

A policy analysis begins with the insuring agreement, insured entity and covered property or interest, then applicable definitions, exclusions, conditions, limits and endorsements. In commercial crime and cyber coverage, relevant mechanisms may include computer fraud, fraudulent transfer instructions, social engineering endorsements, network security liability and breach response expenses. The existence of any of these categories in a market does not establish that a particular contract includes it. An endorsement may limit a subset of events; other policies may provide separate grants. Exact wording and governing law are decisive.

Four appellate records demonstrate why the classification of deceit-induced payments must remain contingent on contract language and factual sequence. In Apache v. Great American (2016), an unpublished Fifth Circuit decision applying a Texas-law prediction, the fraud involved telephone contact, email and subsequent approval processes. The court treated the use of the computer as incidental to payment of genuine invoices to an incorrect beneficiary and rendered judgment for the insurer. The decision illustrates a causation problem, not a universal exclusion of email fraud.

In Medidata v. Federal Insurance (2018), a Second Circuit summary order, manipulation of the email system's displayed sender supported the computer fraud grant. The panel treated the employees' responsive payment actions as insufficient to break proximate causation under New York law. The order is case-specific and limited in precedential status. Its technical facts and policy terms distinguish it from fraud that merely uses a computer as a communication channel.

In American Tooling Center v. Travelers (2018), the Sixth Circuit held that the direct-loss and computer-fraud terms at issue covered the deceit-induced transfers under its analysis of Michigan law. The court declined to add a hacking requirement absent corresponding language. It demonstrates that an employee's act of making a payment is not invariably a superseding cause.

In Ernst & Haas v. Hiscox (2022), the Ninth Circuit reversed dismissal and remanded on computer fraud and funds transfer fraud issues. The 2012 policy's transfer-fraud wording encompassed specified fraudulent instructions communicated to an employee, as alleged. The procedural posture is important: acceptance of pleaded facts on appeal does not establish an ultimate payment or the outcome of later proceedings.

These decisions concern particular crime forms and earlier fraud, not judicial findings about generative AI. Their relevance is analogical and contractual. Synthetic impersonation may change evidence about the identity and content of the instruction; it does not displace the analysis required by the chosen insuring agreement. The following two independent policy specimens supplement the wording reproduced in the opinions. They establish specific contractual designs rather than standard market wording or the frequency of provisions.

Comparative contract specimens

The Zurich Cyber Protect endorsement (Zurich Kotak, n.d., pp. 24–27; product UIN IRDAN152CPLB1295V01202526) distinguishes computer-system or fraudulent-instruction cyber fraud from a good-faith transfer induced by an impostor. Its discovery trigger, direct financial loss requirement, and separate definitions make attribution of a communication to generative AI analytically secondary to the insured event. It stipulates excess positioning relative to other insurance and requires verification before acting on a transfer instruction, through a documented callback or another reasonably secure procedure, with contemporaneous records. Its schedule leaves the limit and deductible to be specified. The specimen therefore establishes the design of a contractual obligation rather than a market-wide numerical coverage rule.

The ECB cyber policy wording (Talbot Underwriting, 2022, pp. 1–3) provides a specific group-policy example: a £50,000 liability limit per covered club, a £10,000 aggregate funds-transfer-fraud and third-party-escrow-theft sublimit within the overall limit, and a £250 excess for those fraud losses. For changed payment instructions, it sets a condition precedent of authentication through previously verified telephone contacts and confirmation. The scheduled amounts pertain to this particular arrangement. They are not estimates of typical corporate cyber limits.

These documents distinguish two sources of coverage variation: the classification of the cause and the procedural eligibility of the payment. A new impersonation medium can increase deception risk while leaving a pre-existing contractual verification condition applicable. The record of verification must consequently be analyzed alongside the actual communication and the policy in force. The broader effectiveness of a control and compliance with a specific contract are separate research variables. These two specimens were examined as indexed PDF text; attempts to retrieve page images failed, so visual comparison with the rendered originals remains outstanding.

Settlement and cross-policy interfaces

A transfer may be economically unrecovered when the claim is notified but partially recovered later. The financial sequence can involve the insured, insurer, recipient bank, sending bank and another policy. Reimbursement and indemnity should be reconciled according to applicable law and recovery terms. Double-counting money returned to the insured would distort the measured economic loss; treating every recovery as a reduction in the insurer's payment without reading the allocation provisions would also be incorrect.

For a hypothetical contract, let X be eligible net loss at the relevant allocation stage, d a per-event retention and L the indemnity sublimit. A simplified payment amount is Y = min(L, max(0, X − d)). Assume X is $130,000, d is $10,000, and L is $100,000. The indemnity is $100,000, leaving $30,000 of eligible loss with the insured. If eligible net loss falls to $80,000 before the calculation, indemnity is $70,000. Actual treatment of expenses, recoveries and retentions depends on the policy. A larger economic loss may lead to the same indemnity under a binding sublimit.

Reporting should distinguish declined, withdrawn, closed without payment and paid claims. An open claim requires separate case estimates and development assumptions. For example, assigning zero to an open transfer investigation just because it is unpaid creates downward bias in incurred-loss measures. Paid claims and reported incidents can have different temporal profiles; actuarial development should use a consistent cohort definition.

Payment reimbursement in comparative context

The UK PSR's PS24/7 reimbursement policy established a maximum reimbursement of £85,000 per in-scope Faster Payments claim effective October 7, 2024. Its APP reimbursement guidance identifies covered customer classes including individuals, microenterprises and charities, and relevant domestic transfer systems. An authorized push payment scam involves deception leading to payment initiation; unauthorized fraud has a different legal treatment. These arrangements allocate losses through payment regulation and do not equate to a cyber insurance contract. Their relevance to insurance is that the insured's residual economic loss and rights against others may depend on the reimbursement regime. Larger corporations, noneligible transfers and other jurisdictions require separate analysis.

The comparison reveals three different legal questions: whether an instruction was authentic and authorized, whether a payment provider has reimbursement obligations, and whether an insurer has a contractual obligation. Answers may diverge. Insurance research should record them independently. The UK policy parameters are tied to the cited effective date; later changes and specific case eligibility should be checked when applying the regime.

Evidence Validation and Sensitivity Analysis

Denominators and selection

The publicly available evidence supports different levels of inference. Human-subject tests measure a narrow outcome under designed exposure. Intelligence reports evaluate capability and trajectory. Police reports document selected or reported events. Insurance claims reports describe a covered portfolio and its reporting and classification practices. Court opinions adjudicate disputes within a defined legal record. Each source contributes to a different link in the explanatory chain. Pooling them as observations from a single statistical population would be methodologically invalid.

The reported AI indicator is subject to both missingness and misclassification. Let q be the actual proportion of AI-assisted events in a defined sample; let s be sensitivity of a classification method and t specificity. If the method is applied uniformly and the population is otherwise comparable, the expected labeled share is r = sq + (1 − t)(1 − q). For example, with q = 0.10, s = 0.70 and t = 0.95, r = 0.115. Among 1,000 events this corresponds in expectation to 70 correctly labeled positives and 45 falsely labeled non-AI events. With no reliable evaluation of s and t in the target claims population, the observed rate does not identify the actual rate. The example illustrates the importance of specificity for a relatively infrequent class; it does not validate a detector.

Unknown technology should remain an explicit category. In a hypothetical review of 200 incident files, assume 20 corroborated AI-assisted events, 100 corroborated non-AI events and 80 indeterminate events. With classifications assumed correct, the sample proportion is bounded from 10% to 50%, depending on the unknown category. The 20% estimate obtained by dividing 20 by 100 would use the wrong denominator; the proportion among known cases is 20/120, approximately 16.7%. If evidence availability is related to the type of fraud, complete-case analysis may be biased. The interval itself addresses missing classification only, not sampling error.

Frequency and severity

For a homogeneous exposure cohort, let N denote annual claim count and Y_j each claim's covered indemnity. Aggregate indemnity is S = sum from j = 1 to N of Y_j. If severity observations are identically distributed and independent of N, E[S] = E[N]E[Y]. Dependence and heterogeneous policy terms require more detailed conditioning. Expected claim cost per exposure is distinct from aggregate volume or total economic loss.

Suppose, purely illustratively, that two scenarios share a population of 1,000 policy-years, with 10 expected covered claims in the first and 15 in the second. If expected indemnity per claim falls from $60,000 to $40,000 due to different conditional loss severity and coverage, each scenario has expected annual aggregate indemnity of $600,000, or $600 per policy-year. The same expected loss can coexist with different uncertainty, operational load and capital requirements. To price a contract, insurers must also account for claim expenses, acquisition and operating expenses, capital and reinsurance costs, uncertainty and profit expectations where applicable.

The payment-limit example earlier demonstrates nonlinearity: E[min(L, max(0, X − d))] generally differs from applying the policy formula to the expected underlying loss. A full loss distribution, with contract terms and recoveries represented consistently, is necessary to estimate tail exposure. Where claims are rare and categories change over time, regularization, credible pooling and bounded stress tests may be more informative than fitting many poorly identified parameters.

Causal inference and alternative explanations

An empirical analysis could compare stable cohorts before and after a specified capability or defensive change, with exposure denominators, policy forms, sector, transaction volume and reporting practices controlled. The period definition and model assumptions would need advance specification. Parallel trends for a difference-in-differences approach, if chosen, require evaluation rather than assertion. Time trends can also reflect more effective recovery, new policy terms, changes to portfolio selection, macroeconomic activity or improved logging and investigation. Insurers with better forensic systems may record more AI-assisted events because identification improves, independent of incidence. Selection into a product or endorsement can generate additional bias.

Validation should preserve out-of-time evaluation, event-level grouping to prevent leakage between related observations, and sensitivity to unknown classifications. For example, an empirical prediction model trained to recognize claims from a single campaign may appear accurate when related files enter both training and test sets. Validation against distinct campaigns and later periods more closely assesses whether the model generalizes. Results should report uncertainty and calibration at the decision threshold relevant to the intended use. No predictive model is fitted in this research.

Reserve and accumulation implications

Reporting lag and recovery lag create distinct development problems. Claims may be notified promptly while disputes over coverage, recovery and responsibility remain unresolved. Aggregating immature and mature claim cohorts can lead to misleading severity trends. Insurers should reconcile paid amounts, expected future payments, allocated adjustment expenses and recovery rights using their applicable accounting and reserving standards. A change in reporting labels should be bridged historically rather than retrospectively manufacturing an AI time series.

Stress tests can vary exposure intensity, consequential-action probability, net severity, policy limits, reliance on third parties and the probability of simultaneous losses. Results depend on the joint distribution rather than a single scenario mean. A scenario that assumes large losses and simultaneous coverage across several policy classes may exceed realistic contract and process conditions; the rationale for each joint assumption should be recorded. The examples in this dossier are arithmetic sensitivities, not forecasts of capital or insolvency.

Integrated Scientific Synthesis and Research Agenda

Information structure for claims research

The durable contribution of the research is the integration of technology attribution, financial event classification, contract interpretation and economic development. A feasible research record begins with an event identifier, related campaign identifier if corroborated, affected insured entity and policy period. It records channel, impersonated role, credential compromise status, AI involvement and confidence, action taken, financial event type, gross transfer, recovery, expenses, policy grant and adjudication. Additional fields distinguish report date from event date and settlement date. Monetary values are stored with currencies and consistent definitions. This is a suggested data structure derived from the analytical distinctions, rather than a claim of theoretical novelty.

The AI field should support documented evidence and uncertainty without making the classification a precondition for standard claim handling unless the contract requires it. A file can have a well-established fraudulent transaction and uncertain technology attribution. Processing the coverage question can proceed on its own evidence. This creates a more reliable research dataset and avoids imposing unsupported forensic conclusions on policyholders.

Use controlled vocabularies at the stage of data collection, while preserving original narrative and the source's own category definitions. A single primary claim label may support reporting, but ancillary events must remain discoverable. For example, a mailbox compromise can precede a transfer and a privacy investigation. The primary loss classification should not erase secondary costs. Conversely, recording three related labels as three claims inflates frequency. Claims data should link multiple cost components to a unique claim, and policy-level and incident-level aggregations should be reproducible.

Contract clarity and process ownership

Underwriting, claims, risk engineering and legal personnel require shared definitions without conflating their functions. Underwriting establishes prospective terms; risk engineering evaluates processes; claims investigates realized events; legal analysis interprets coverage and recovery rights. Clear communication about sublimits and insured interests supports informed decisions, while an ambiguous application question about AI can induce inconsistent classification. Underwriting information should be proportionate, verifiable and relevant to residual loss rather than merely expressing generalized concern about new technology.

Post-event remediation can improve observability and reduce repeat loss, but an organization's adoption of controls after an incident creates a timing issue in research. Current controls should not be used as though present at the historical loss date. Store the installation and effective operational dates and distinguish evidence supplied retrospectively. Studying recovery also requires consistent treatment of denied claims, because the absence of an insurer payment is not proof that the underlying economic loss did not occur.

Research priorities

Future research would benefit from collaborative, anonymized event-and-claim records using a common taxonomy with protected access, purpose limitation and auditability. Priority questions are the effect of generative assistance conditional on exposure and controls; the marginal effectiveness of independent transaction verification; the distribution of recovery times; the consequences of alternative endorsement wording; and the extent of campaign dependence across insured organizations. Each requires a clearly identified denominator and independently recorded outcomes. A longitudinal design that preserves changes in policy terms and forensic capacity would address important confounding factors.

A second priority is comparative law beyond the selected reported examples. Access to a broader representative sample of contemporary endorsements, applicable local legal instruments and adjudicated claims would allow a more representative comparison of payment fraud and insurance interfaces. The present judgments provide valuable primary evidence about specific wordings, but their doctrinal reach is bounded. The public sources give no basis to estimate a world aggregate of AI-attributable paid social engineering claims.

A third priority concerns the ethics of defensive monitoring. More detailed attribution data may be useful for fraud investigation but can expose personal information, employee communications and sensitive financial records. Use purpose-limited retention, access controls, lawful sharing and review of automated classifications. Risk transfer should support resilience without creating incentives to collect irrelevant personal characteristics or assume individual culpability solely from victimization. The technical possibility of monitoring does not establish its legal permissibility or proportionality.

The scientific synthesis thus places AI assistance in a broader loss process already familiar to insurance science. The significance of generative capabilities is their effect on communication scale and credibility, conditional on organizational workflows and the surrounding defensive environment. The ultimate insurance consequence is established through loss realization, recoveries and the particular policy. These relationships explain why plausible changes in attack capability may coexist with rising, flat or falling covered loss in different portfolios.

Conclusion

Generative AI phishing is a material subject for cyber insurance research because it can alter the frequency and character of deceptive communications and complicate the authentication of apparent authority. Experimental evidence supports the capacity of personalized generated messages to elicit interaction, while official case evidence establishes that synthetic impersonation has been associated with substantial economic losses. Public complaints and insurer summaries demonstrate the continuing importance of BEC and funds transfer fraud. These sources measure different outcomes and cannot identify a global AI-specific insurance loss total.

The principal insurance findings are that the initial communication, realized financial loss, evidence of AI use, policy grant and subsequent recovery require separate analysis; that employee participation in payment initiation does not by itself settle coverage; and that contract wording and jurisdiction govern classification. Frequency, severity, coverage limits and dependence must enter actuarial analysis through consistent definitions. Defensive investment can address a consequential decision or recovery stage even when deceptive communication remains difficult to identify.

This dossier contributes an original insurance-centered synthesis of existing documentary and research evidence and transparent numerical illustrations, without claiming a new theory or an empirically estimated AI loss multiplier. The evidence supports the research conclusions at their stated scope. Further validation needs linked, mature and consistently classified claims data, full modern policy wordings, and broader jurisdictional coverage. The enduring scientific question is how changing capabilities are transmitted through institutional decisions into economic loss, and how accurately insurance transfers that residual loss.

References

  • American Tooling Center, Inc. v. Travelers Casualty & Surety Co. of America, 895 F.3d 455 (6th Cir. July 13, 2018). American Tooling opinion

  • Apache Corp. v. Great American Insurance Co., No. 15-20499, 662 F. App’x 252 (5th Cir. Oct. 18, 2016) (unpublished). Apache opinion

  • Coalition. (2026). 2026 Cyber Claims Report: Key Findings (calendar year 2025). Cyber Claims Report

  • Czybik, S., Kouam, A. J., Heubl, P., Nold, J. M., & Rieck, K. (2026). A large-scale study of personalized phishing using large language models. 35th USENIX Security Symposium. Personalized phishing study

  • Ernst & Haas Management Co., Inc. v. Hiscox, Inc., 23 F.4th 1195 (9th Cir. Jan. 26, 2022). Ernst & Haas opinion

  • European Union Agency for Cybersecurity. (2026, September 22). Exploring the evolution of the cyber threat landscape: How dependencies weaken our digital resilience. Threat landscape summary

  • Federal Bureau of Investigation, Internet Crime Complaint Center. (2026). 2025 IC3 Annual Report, pp. 26, 39–40. IC3 Annual Report

  • Government of the Hong Kong SAR, Security Bureau. (2024, June 26). LCQ9: Combating frauds involving deepfake. Legislative Council written reply. Legislative Council reply

  • Heiding, F., Lermen, S., Kao, A., Schneier, B., & Vishwanath, A. (2024). Evaluating large language models’ capability to launch fully automated spear phishing campaigns: Validated on human subjects. arXiv:2412.00586. Spear phishing study

  • Ho, G., Cidon, A., Gavish, L., Schweighauser, M., Paxson, V., Savage, S., Voelker, G. M., & Wagner, D. (2019). Detecting and characterizing lateral phishing at scale. 28th USENIX Security Symposium, 1273–1290. Lateral phishing study

  • Medidata Solutions Inc. v. Federal Insurance Co., 729 F. App’x 117 (2d Cir. July 6, 2018) (summary order). Medidata opinion

  • National Cyber Security Centre. (2024, January). The near-term impact of AI on the cyber threat. Near-term threat assessment

  • National Cyber Security Centre. (2025). Impact of AI on cyber threat from now to 2027. 2027 threat assessment

  • National Institute of Standards and Technology. (2025). Digital Identity Guidelines: Authentication and Authenticator Management (SP 800-63B-4), Authenticators, Phishing Resistance. Phishing resistance

  • National Institute of Standards and Technology. (n.d.). Cybersecurity Framework: Frequently Asked Questions (CSF 2.0). CSF 2.0 FAQ

  • Payment Systems Regulator. (2024). PS24/7: Faster Payments APP scams reimbursement requirement: Confirming the maximum level of reimbursement. PS24/7 policy statement

  • Payment Systems Regulator. (n.d.). APP fraud reimbursement protections. APP fraud protections

  • Steves, M., Greene, K., & Theofanos, M. (2020). Categorizing human phishing difficulty: A Phish Scale. Journal of Cybersecurity, 6(1), tyaa009. Phish Scale

  • Zurich Kotak General Insurance Company (India) Limited. (n.d.). Zurich Cyber Protect: Add-on/Endorsement Wordings. UIN IRDAN152CPLB1295V01202526, pp. 24–27; endorsement UIN IRDAN152CPLB1295V01202526/A1310V01202526. Cyber Protect endorsement

  • Talbot Underwriting Ltd. (2022). ECB Cyber Policy Wording. Master policy UMR B1740PGL210088000, pp. 1–3. Howden-hosted PDF (April 2022). ECB policy wording