RESEARCH DOSSIER
Vehicle Cyber-Physical Failures and Motor Liability Accumulation
Abstract. Connected vehicles join mechanical control, embedded computation, software maintenance, external communications, human operation, infrastructure, and insurance contracts in a single loss-producing environment. A cyber-physical failure becomes an insurable motor-liability exposure when a technical disturbance contributes to a legally relevant accident, causes compensable harm, establishes a liability or statutory compensation obligation, and falls within an operative insurance or indemnity arrangement. Common software, hardware, cloud services, connectivity, credentials, update pipelines, map data, and mobility infrastructure can couple events across vehicles that appear independent in conventional motor rating datasets. This monograph examines the complete sequence from technical vulnerability through physical consequence, causal proof, legal allocation, insurance coverage, portfolio aggregation, reserving, reinsurance, and capital. It develops a documented comparison of automotive experiments and safety notices; the EU, United Kingdom, and United States legal contexts; international cybersecurity and supervisory standards; and cyber-insurability scholarship. Transparent synthetic portfolio calculations show that a small marginal accident probability can coexist with severe common-event concentration, and that identical expected losses permit very different aggregate tails. All numerical stress scenarios are explicitly hypothetical; documented recalls and laboratory exploits are not treated as observed mass accidents. The original scholarly contribution is a source-linked, cross-disciplinary analytical synthesis and reproducible interpretation of established probability identities, rather than a newly named theory, estimator, or doctrine. The resulting approach separates technical exposure, causal attribution, legal compensation, contractual payment, and ultimate retained insurer loss and identifies the information required to validate a credible motor cyber-physical accumulation model.
Keywords: Connected vehicles; cyber-physical systems; automotive cybersecurity; motor liability; cyber insurance; common-mode failures; correlated claims; motor catastrophe; vehicle automation; software updates; liability allocation; accident causation; product liability; non-affirmative cyber; actuarial aggregation; reinsurance; solvency; risk concentration; claims evidence.
Introduction
A vehicle can be exposed to a common technical cause while suffering an individually localized collision. A defect in one supplier component, a compromised fleet administration service, a flawed update, or a maliciously manipulated information feed may affect vehicles in different places and at different times. The resulting legal claims are often administered as individual motor accidents. For a motor insurer, however, the loss distribution may depend on a shared technical event rather than the independence assumptions that make geographical diversification effective against many ordinary collisions. This is the central problem of vehicle cyber-physical failures and motor liability accumulation.
The underlying safety problem is consequential on a global scale. The World Health Organization's Global Status Report on Road Safety 2023 estimates approximately 1.19 million annual road traffic deaths. The report describes road safety generally; it supplies neither a count of cyber-induced collisions nor a denominator for cyber-exposed vehicle operation. Cyber-physical accumulation therefore requires its own cause-specific evidence and cannot be calculated by relabeling the aggregate road-traffic burden.
The technical plausibility of remote interference has been established in controlled research. Koscher and colleagues (2010) investigated control pathways within a studied vehicle. Checkoway and colleagues (2011) extended analysis to external attack surfaces and remote pathways. These are foundational demonstrations of mechanisms and preconditions, not estimates of nationwide malicious crash occurrence. The distinction between observed technical capability and realized insured injury runs through every subsequent chapter.
Three research questions organize the dossier. First, how does an electrical, computational, or communication disturbance become an identified accident-causing event? Second, how can shared dependencies create many incidents and claims, and what levels of dependence matter to an insurer? Third, how do different legal and contractual mechanisms turn those incidents into compensation, reserves, gross insured obligations, net losses, and reinsurance recoveries? The analysis applies to passenger vehicles, commercial fleets, automated mobility systems, and shared infrastructure, with the observation unit stated for each numerical example.
The evidentiary method combines original technical studies, official recall documentation, standards catalog records, laws in their stated versions, insurance supervisory materials, and academic cyber-risk research. Legislative descriptions are tied to jurisdictions and dates. The statistical examples use explicitly assigned assumptions and algebraic identities, with no claim of fitting incident data or insurer portfolios. Material ambiguities concern the origin of particular accidents, the legal scope of indemnity, the concentration of supplier dependencies, and the breadth of event aggregation language in individual reinsurance contracts. The research cutoff is November 16, 2026.
Scientific Definitions and Primary Foundations
From computer disturbance to insured loss
A cyber-physical vehicle is a coupled system in which software, sensors, communications, actuators, mechanical components, human drivers, and external operating conditions can affect physical motion. A cyber-physical failure is a deviation of this coupled system from its intended safe functioning attributable to the digital or electronically mediated part of its operation. The immediate source can be malicious interference, defective code, corrupt or delayed input, software configuration, a failed update, a hardware or sensor fault, or loss of a needed external service. Those causes must be distinguished in investigation and insurance classification: a cyber-enabled failure is not automatically a criminal cyberattack, and a cyberattack does not necessarily generate physical damage.
The complete insurance analysis follows an ordered chain: technical source; exploit or malfunction; safety-critical control effect; accident or hazardous state; claimant harm; legal cause and responsibility; compensation entitlement; insured event and coverage; claims reporting; reserve and settlement; reinsurance recovery; and residual financial loss. Each arrow represents a proposition that may require different evidence. A software event can remain a mere vulnerability, yield a successful control intervention without an accident, cause an accident without legal liability on a particular defendant, or generate liability that a particular policy does not respond to. Insurance accumulation concerns the last stages of this chain while retaining their dependence on earlier causal conditions.
The unit of observation is fundamental. A vulnerable vehicle is an exposure, not a claim. A software episode is a technical event, not necessarily an insurance occurrence. One collision may injure several persons, damage several vehicles, and trigger several policies. A claimant may submit multiple bills against one accident; a manufacturer may incur one product claim after a motor insurer has already compensated the victim; an insurer may report one claim number but several separately insured coverage payments. Confusing these units can produce apparent frequency increases or duplicated insurer losses. Every model must therefore declare which count it represents.
Security engineering and safety governance
The UNECE cybersecurity regulation, UN R155 establishes a vehicle-type-approval framework centered on cybersecurity management, while UN R156 on software updates addresses update-management systems. Their legal reach varies with the type-approval system and parties applying the regulations; neither document promises a zero-failure vehicle. The related ISO/SAE 21434:2021 standard concerns lifecycle automotive cybersecurity engineering. Here the public ISO catalogue verifies identity and scope; detailed conformity judgments require the licensed normative text and vehicle-specific audit evidence.
The U.S. National Highway Traffic Safety Administration's 2022 vehicle cybersecurity guidance explicitly describes voluntary, nonbinding best practices. It shows the relevance of continuous organizational security and risk management without transforming guidance into a nationwide compulsory technical guarantee. The ENISA–JRC autonomous-driving report adds learning-system and data-validation issues: an adversarially changed sensor input can alter system perception, while model updates and dependencies create lifecycle questions. Affected perception is still a technical condition to be connected to an actual accident before legal or actuarial consequences are quantified.
Security and safety management intersect at several measurable controls: the set of software versions in circulation, the distribution of signing keys, the feasibility of rollback and isolation, the mechanism of remote feature activation, the ability to detect unauthorized commands, and the quality of post-incident logs. A completed type approval, penetration test, or vulnerability disclosure is evidence about a specified process or configuration. It cannot be substituted for an estimated distribution of policyholder injury or for a verified claim payment. A robust research design uses engineering tests to bound plausible mechanisms and uses accident and insurance evidence to investigate their consequences.
The difference between risk, liability, coverage, and accumulation
In this dossier, technical exposure denotes the intersection of a vulnerable system and conditions under which the vulnerability could affect a vehicle. Physical risk denotes a probability distribution of adverse physical outcomes conditional on exposure, traffic, operator behavior, and defenses. Legal liability denotes an applicable obligation to compensate under a named legal rule and the case facts. Insurance coverage denotes the scope, attachment, exclusions, conditions, limits and priorities of the relevant contract or statutory insurance framework. Accumulation denotes the joint distribution of obligations arising from one or several dependence-generating causes across insured units, contracts, insurer entities and reinsurance layers.
U.S. consumer motor coverage distinguishes bodily-injury and property-damage liability and other forms of protection, as shown by the NAIC auto insurance explanation. Contractual language and state law establish the actual allocation. It is accordingly misleading to assume that a cyber-originated crash will be reimbursed by a separately purchased cyber policy, excluded from motor liability, or transferred wholesale to the vehicle manufacturer. Each conclusion requires a particular liability pathway and policy wording.
Two dimensions must remain explicit: horizontal correlation across vehicles and vertical accumulation across coverage and financial layers. One software flaw could create accident claims against several auto insurers, third-party product liability against a manufacturer, technology errors-and-omissions claims against a vendor, fleet business interruption, and losses on reinsurance treaties. Some entries represent the same ultimate victim harm at different points of payment or recovery. Summing all gross figures without accounting for subrogation, contribution and offsets double-counts the economic loss. Conversely, analyzing only the manufacturer omits claims first paid under motor insurance.
Cyber-Physical Malfunction Events
Technical entry routes and control consequences
The original 2010 experimental vehicle-security study showed how access to electronic control units and internal networks could alter safety-relevant functions in its experimental environment. The later 2011 attack-surface experiments described pathways through interfaces such as diagnostics, media, wireless services, and cellular connectivity. These papers establish technological feasibility conditional on particular vehicle architectures, access routes and exploit capabilities. Their results are neither statements about every vehicle on the road nor physical-accident frequency estimates.
The Miller–Valasek 2015 white paper investigated a remotely accessible, unaltered passenger vehicle and illustrated the path from an external communication interface through internal networks to physical-system effects. The related official NHTSA recall 15V-461, reported July 23 and acknowledged July 24, 2015, identified a potential affected population of 1,400,000 vehicles with radio software vulnerabilities permitting unauthorized access to certain networked control systems. The notice states increased crash risk as a possible consequence; it does not substantiate 1.4 million crashes, claims, or injured persons. The recall documents the substantial size of a common vulnerable population and the relevance of preventive remedy, not realized loss severity.
The resulting mechanism can be represented by five stages. First, a reachable and vulnerable interface exists. Second, the attacker or unintended disturbance obtains an effective pathway into a safety-relevant function. Third, a command, corrupted datum, software state, or control failure changes vehicle behavior. Fourth, the traffic and human-control context permits a collision or other harm rather than recovery. Fifth, the event and resulting injury satisfy fact-specific causal and legal tests. Each stage has a different failure probability, which will vary across model years, software revisions, vehicle states, geographic conditions and defensive configurations.
Software update failures and changing risk cohorts
Software changes create time-dependent risk cohorts. Vehicles may receive the same faulty package at different times; vulnerable versions can remain on offline vehicles; a remedy may disable one remote entry route while not eliminating every local route; rollback may itself be unsafe in some configurations. UN R156's update-management requirements identify the importance of managed updating, whereas ISO/SAE 21434 lifecycle scope covers risks from development through operation and decommissioning. For accumulation, an insurer needs version-level installed-base data and time stamps, not merely manufacturer and model labels.
A software defect that is distributed to many cars is an important common cause even in the absence of hostile interference. The causal path might be a braking-calibration error, unanticipated interaction among control features, or inconsistent configuration after update. A maliciously altered package may have similar physical consequences while raising distinct forensic and liability issues. Classifying both as a common cyber-physical origin is useful for exposure aggregation, provided legal and policy analyses continue to distinguish defective manufacture, inadvertent error, unauthorized tampering and later failure to remedy a known defect.
Update penetration creates a moving denominator. Suppose an identified software version is installed on 80,000 vehicles during a week, but only 12,000 are in the insurer's portfolio and only 3,000 operate in the relevant geographic or control-mode context. The first number describes a manufacturer population, the second the insurer's contractual exposure, and the third a conditional risk set for a specific mechanism. These three counts are illustrative. Without comparable timestamps and unique vehicle identifiers, multiplying a model-level manufacturer recall count by an insurer's overall motor market share may materially overstate or understate the relevant insured fleet.
Infrastructure, data integrity, and automation
Vehicle-to-infrastructure messages, mapping and positioning data, cloud fleet administration, route optimization, remote dispatch, credential services, and charging or telecommunications networks create additional dependence paths. Some failures cause loss of a convenience service; others might corrupt a safety-relevant command or undermine a required operational decision. The connection must be documented at the vehicle's actual operational design domain, rather than inferred from the mere presence of a network connection. An automated vehicle can be designed to enter a safe fallback upon signal loss; a common communications outage does not necessarily result in simultaneous crashes.
The ENISA–JRC analysis of autonomous-driving AI identifies risks associated with adversarial data, model performance and lifecycle changes. Those risks require separate exposure definitions for sensor perception, planning, control, and communications. Automation levels are also material. NHTSA's Standing General Order reporting framework distinguishes automated driving systems from Level 2 assistance, which retains a continuously engaged human driver. A malfunction in an assisted-driving feature therefore raises different immediate human-control questions from malfunction during truly automated operation.
Evidence required at the accident level
A defensible incident file should preserve vehicle identity, model year, hardware revision, software and firmware versions, diagnostic logs, event-data recorder records where available, relevant sensor inputs, time synchronization, update and credential history, network and cloud logs, safety-state transitions, command provenance, driver alerts and responses, collision reconstruction, and physical inspection. Authentication and chain of custody must be recorded. A successful remote-control demonstration is a mechanism hypothesis; a real claim requires traceable evidence that the relevant mechanism occurred and contributed to the crash under the governing causal standard.
Attribution is especially difficult when competing explanations remain possible: unsafe driver behavior, road conditions, maintenance, sensor obstruction, unrelated equipment failure, adversarial tampering, and software error. Investigators should compare a hypothesized technical cause against alternatives, establish the temporal sequence and operational opportunity, and preserve contradictory evidence. Where logs have gaps or can be overwritten, the evidentiary record should state what cannot be known. Absence of a recovered malware signature does not prove a benign mechanical origin; the presence of a software vulnerability does not prove malicious use in a particular crash.
The claim-level chronology should separate the date of technical compromise, manifestation in physical motion, collision, first report to an insurer, proof of liability, reserve revision, and cash payment. Multiple accidents can share one upstream cause yet create different dates of loss, legal jurisdictions, and coverage periods. Accumulation decisions made solely from a vendor's incident date may conflict with the contractual dates that govern each motor policy and reinsurance contract.
Common Systems and Multi-Vehicle Losses
Dependence-producing infrastructure
The economically decisive distinction is between sharing a vehicle characteristic and sharing an active failure cause. Two vehicles may use the same electronics vendor without sharing a vulnerability capable of affecting braking. Conversely, vehicles sold by different manufacturers may depend on one underlying service, library or credential authority. Common failure therefore must be specified as an actual dependency edge: software package or component; vulnerable version; exploitable condition; common remote service; common adversarial campaign; or common environmental and network condition. Fleet size by brand alone is an incomplete exposure measure.
Dependence occurs at several nested scales. A single vehicle can experience simultaneous subsystem failures; a commercial fleet can suffer a common software or dispatch outage; geographically separated fleets can share a technology vendor; several motor insurers can insure different parts of the same vulnerable installed base; and a reinsurer may accept multiple cedents' correlated losses. This nested observation resembles the distinction between within-enterprise and across-enterprise dependence in Böhme and Kataria (2006). Their cyber-insurance paper concerns information systems generally. Applying its dependence insight to motor physical injury is a transparent analogy and requires motor-specific validation.
Cross-vehicle correlation may be generated by a single latent event even when accidents remain conditionally independent after that event. For example, if a compromised service activates a risky operating mode across a fleet, particular crashes can still depend separately on speed, nearby traffic, location, human override and environmental conditions. This distinction matters because a common technical failure does not imply that every affected vehicle crashes. At the other extreme, a configuration change that simultaneously disables a safety function may be followed by concentrated incidents in one geographic traffic environment. Portfolio models should allow both forms and test the dependence they imply.
The cyber-insurability literature provides context without direct motor calibration. Biener, Eling, and Wirfs (2015) analyzed 994 cyber-loss cases in an operational-risk dataset, emphasizing sparse data and linked losses as obstacles to insurability. That database should not be reinterpreted as measured motor collision severity or a frequency of automotive cyber casualties. Its relevance here is methodological: ordinary independent-loss methods can miss joint shocks when exposures share technology and threats.
A transparent two-stage portfolio stress calculation
Consider a hypothetical one-year portfolio of 10,000 separately insured vehicles. Define a common disturbance indicator Z that equals one with probability q = 0.01. Conditional on Z = 1, each vehicle has an independent probability p = 0.02 of experiencing one covered liability-producing physical accident during the modeled disturbance horizon; conditional on Z = 0, this particular modeled mechanism causes no accidents. Every resulting claim has a fixed insured gross payment of $250,000, with no deductibles, sublimits, litigation costs, recoveries or inflation. These assumptions are deliberately strong and invented only to expose the arithmetic of shared vulnerability; they are not forecasts or estimates from accidents, recalls, or insurer data.
Let Y_i be independent Bernoulli(p) variables under the disturbance, and define X_i = ZY_i, the indicator of a modeled insured accident for vehicle i. The mechanism-specific number of claims is C = sum of X_i across N vehicles, and the associated gross insurer payment is G = sC, where s is $250,000. Direct conditioning yields E[C] = Nqp and E[G] = sNqp. For distinct vehicles, Cov(X_i, X_j) = q(1−q)p², and the count variance is
Var(C) = Nqp(1−p) + N²q(1−q)p².
The first term is idiosyncratic variation within a disturbance; the second is the uncertainty about whether a common disturbance occurs at all. These identities follow directly from the law of total variance and the stated independence assumptions. They do not constitute a new actuarial statistical model or a novel estimator.
For N = 10,000, q = 0.01 and p = 0.02, the expected number of mechanism-specific claims is 2, yielding expected annual gross loss of $500,000. Conditional on the disturbance, the expected count is 200 and expected gross payment is $50 million. The variance of C equals 397.96 and the standard deviation of G is approximately $4.987 million. A model treating the 10,000 vehicles as independent Bernoulli risks with the same marginal event probability qp = 0.0002 retains the $500,000 expectation, but gives count variance 1.9996 and gross-loss standard deviation approximately $353,518. The difference arises from shared event risk, not different mean claim costs.
This exercise makes a practical distinction between annual expected cost and tail capital need. The common-shock construction has approximately a 1% annual chance of an adverse event capable of producing a conditional expected aggregate of $50 million. It does not follow that the 99th percentile of loss must equal $50 million: discrete mixture quantiles require the exact chosen convention and conditional binomial distribution. Stress severity, layer exhaustion and risk margins therefore need distributional calculations or simulation rather than an appeal to the mean alone.
Equal expected losses, different concentration
Hold N = 10,000, s = $250,000 and qp = 0.0002 fixed. Three distinct hypothetical parameter combinations are: a frequent, smaller disturbance q = 0.02 and p = 0.01; the baseline q = 0.01 and p = 0.02; and a rarer, more severe disturbance q = 0.002 and p = 0.10. All yield expected count 2 and annual gross expected loss $500,000. Their conditional expected claim counts are respectively 100, 200 and 1,000; their conditional expected gross payments are $25 million, $50 million and $250 million. They therefore create radically different event-scale attachment and exhaustion consequences for a reinsurer, even with identical mean motor loss.
The corresponding count variances, derived from the same formula, are 197.98, 397.96 and 1997.8, respectively. This ranking is driven by the magnitude of a common loss when the adverse state occurs and demonstrates why modeling only observed annual loss frequency can conceal event tail uncertainty. The exercise is sensitive to the explicit assumption that conditional accidents are independently generated with constant p and constant severity. If physical crashes cluster through traffic interactions, if claim severities correlate within one accident, or if a common disturbance persists over multiple days, additional dependence must be represented. The displayed numbers should not be imported into actual premiums or capital calculations.
A market-wide example shows another layer. Assume the 10,000 vehicles are divided evenly across five insurers, each writing 2,000 of the exposures to the same shared component. For the baseline disturbance, each insurer has expected conditional losses of $10 million, while the five portfolios together have conditional expected losses of $50 million. The insurers may appear diversified from one another by customer and region yet share the same technology-source risk. If the five also purchase catastrophe reinsurance from one group, the reinsurer's accumulation can reassemble their dependence. A counterparty-level aggregation must avoid counting a cedent's gross claim twice as both primary and reinsured economic loss.
Which kind of distribution is being modeled?
A count model may operate at four levels: the number of disturbance episodes, vehicles experiencing hazardous effects per episode, physical accidents per affected vehicle, and claims or indemnity payments per accident. A hierarchical model may use a frequency process for shared episodes, vulnerability-conditioned vehicle impact probabilities, an accident conversion process, and claim severity by injury and property damage. Such a construction is a conventional scenario-modeling architecture assembled from recognized statistical components. Its parameters require observed or engineered evidence and should be reported separately rather than treated as one fitted cyber crash rate.
Common-shock dependence can be introduced using an explicit event indicator, random effects, a stochastic intensity, correlated Bernoulli exposures, shared components in a dependency graph, or a simulation of system and traffic states. No one mathematical family has intrinsic authority over the others. A Poisson model for common incidents could be reasonable for a stable threat process, while clustered cyber campaigns and changing defenses may violate stationary Poisson assumptions. A copula can encode dependence given marginal distributions, but may hide the physical source of a common disturbance. A dependency graph clarifies exposed pathways, yet graph topology alone gives no validated probabilities. The appropriate model follows the question and the available validation evidence.
Tail risk requires attention to concentration, timing and conditional severity. A region-wide road event can create many victims and claims within a single collision chain; software deployment creates a broad but possibly temporally dispersed cohort; shared sensor data may influence automated vehicles operating in similar environments. These cases can generate equal aggregate payments with different occurrence definitions. A model must therefore retain a technical event identifier, individual accident identifier, claimant identifier and contract-period identifier, with an auditable mapping between them.
Accumulation by peril, contract, and reinsurer
For insurer k and claim j, let the net eligible contract payment be P_kj after applying the operative motor coverage terms and legal entitlement. The insurer's gross portfolio obligation is the sum of its P_kj, without assuming that all claims involve one insurance occurrence. To evaluate an occurrence excess-of-loss treaty, the contractual aggregation rule must first map individual claims into occurrences e, producing covered occurrence totals H_ke. A stylized treaty payment with attachment A and layer limit L is R_ke = min(max(H_ke − A, 0), L), conditional on all other treaty wording and eligibility conditions being satisfied. The cedent's net retained amount is then gross eligible payment less actual collectible reinsurance recoveries, subject to reinstatements, exclusions and credit risk.
A common software defect does not itself establish that every resulting accident constitutes one reinsurance occurrence. Wording may refer to one accident, originating cause, event, series of losses, geographic and temporal restrictions, or other negotiated conditions. Separate liability treaties, motor excess layers, product liability placements, cyber contracts and facultative covers can define the unit differently. Reinsurance aggregation is therefore partly a contract-interpretation problem, and the modeling file must preserve alternative event-clustering assumptions rather than assigning a universal cyber event rule.
Using the baseline conditional expected gross of $50 million purely as a deterministic stress figure, a hypothetical reinsurance layer of $10 million excess of $20 million would pay $10 million if all underlying eligible losses were legally grouped as one covered occurrence and the cedent met every treaty condition. If the individual accidents were separately treated and none exceeded the $20 million attachment, that same stylized per-occurrence layer might pay zero. This is an illustration of event-definition sensitivity, not a legal finding about any actual treaty. Treating recoveries as certain before resolving occurrence interpretation could overstate net protection.
Liability Allocation and Insurance Aggregation
The chain of possible legal responsibility
A cyber-physical vehicle accident can implicate the driver, vehicle keeper or owner, transport operator, automotive manufacturer, component supplier, software developer, cloud service provider, infrastructure operator, fleet administrator, and an intentional attacker. Their potential responsibilities arise under different legal bases: driver negligence, statutory motor compensation, product defect liability, contractual undertaking, employer or operator duty, professional negligence, and intentionally unlawful conduct. The defendant most directly responsible for the technical defect is not necessarily the institution that first pays an injured road user.
Legal causation requires a sequence of factual and normative judgments. Investigators first consider whether the alleged disturbance was present and contributed to the vehicle behavior; the applicable law then establishes duty, defect, fault, statutory responsibility, causation standards, defenses, recoverable damage, limitation rules and procedural routes. Technical provenance is essential but does not alone resolve legal responsibility. A negligent driver who fails to maintain situational awareness, a faulty emergency takeover process, and a defective safety feature may coexist within one collision. An insurer must analyze the claim against the actual governing law and policy, not against a general label such as 'hacked car.'
Recovery rights are distinct from primary compensation. Where a motor insurer compensates victims and then seeks contribution or indemnification from a manufacturer or other liable party, the insurer's gross claim and later recovery should be kept as separate cash-flow entries. Litigation expenses, defense obligations, multiple liability defendants and delay can prevent the net economic amount from matching the engineering allocation of blame. A single physical accident may produce a timely motor claim and a long-tailed disputed products action.
European Union: victim protection and product claims
The consolidated EU Motor Insurance Directive harmonizes central aspects of compulsory insurance against civil liability arising from vehicle use. Directive (EU) 2021/2118 amended the framework, including provisions addressing injured parties and insurer insolvency. The consolidated text defines 'use of a vehicle' by reference to its function as transport at the time of accident, with specified coverage rules. National transposition and claims law determine the concrete route of redress. A cyber origin does not by itself remove a traffic victim from the logic of compulsory motor protection.
The EU's General Safety Regulation 2019/2144 establishes vehicle safety and type-approval obligations, including technical dimensions of automated driving. Compliance findings, however, should be assessed separately from the existence of a private civil cause of action or the amount of a statutory compensation entitlement. Harmonized technical requirements and civil-liability instruments address related but distinct questions, and safety compliance alone is not an actuarial estimate of future claim costs.
The revised Product Liability Directive (EU) 2024/2853 recognizes software as a product within its relevant rules and addresses integrated digital dependencies. Its corrected Article 2(1) applies the new directive to products placed on the market or put into service after December 8, 2026; as of the November 16, 2026 research cutoff that date remains in the future. A rule governing prospective defective products should not be assigned retroactively to a 2015 recall or to vehicles governed by a different applicable regime. Its relationship to compulsory victim compensation also requires distinct analyses of the specific defendant, relevant product, country and loss facts.
The prospective legal structure makes two concurrent financial pathways plausible: a victim may be compensated through a motor liability mechanism while an insurer or victim may seek further recovery from a product manufacturer or technology supplier under applicable law. The eventual distribution of payments depends on rights of recourse, proof of defect, loss causation and available defendant assets. Counting the victim's injury once as a motor payment and again as a recovered manufacturer payment would confuse a sequence of transfers with two independent underlying physical harms.
United Kingdom: first-instance motor insurance and automated operation
Section 2 of the Automated and Electric Vehicles Act 2018 establishes specified insurer liability where an accident is caused by an automated vehicle while driving itself on a road or other public place in Great Britain, the vehicle is insured, and a person suffers qualifying damage. The Act also addresses qualifying cases involving certain uninsured exempt owners. The statutory scheme describes the insurer's initial liability within its conditions, including a defined property-damage limit, and does not erase other persons' potential responsibility. Other provisions address contributory negligence, unauthorized software changes, safety-critical update failures and recovery against responsible persons.
The Automated Vehicles Act 2024 develops a more detailed authorization and responsibility framework for automated driving. Its enactment is distinct from commencement of each relevant provision, and the official revised legislation warns that changes may have different effective dates. Any litigation assessment must check the provision actually in force on the accident date, whether a vehicle and feature meet the statutory authorization criteria, and who had the legally prescribed operational role. A vehicle marketed with driver assistance should not be assumed to satisfy the requirements for the particular automated-driving liability regime.
The United Kingdom example illustrates why insurance coverage cannot be deduced from a cybersecurity assessment. The applicable automated-driving liability route, a conventional driver-liability route, and a product-defect recourse action can be analytically connected yet legally distinct. Their evidentiary thresholds and timing can differ, and the recoveries may arrive years after the first motor claim. Appropriate reserving must incorporate litigation uncertainty and expected recoveries separately.
United States: decentralized motor and products rules
U.S. motor liability remains materially shaped by state-specific insurance requirements, liability laws, policy wording and financial responsibility arrangements. The NAIC's consumer motor guidance provides a basic taxonomy of bodily-injury and property-damage liability coverage, but does not establish a uniform national decision rule for cyber-originated accidents. State-specific negligence, products liability, statutory limits, exclusions and subrogation can change the result. Federal vehicle safety oversight and recall obligations operate in a different legal domain from a private insurer's claim and indemnity decision.
The NHTSA 2015 radio-security recall is an official record of a technical hazard and remedial response. It does not establish that a particular victim has sustained an insured loss, that each recalled vehicle was injured, or that a party is liable under a specified state tort rule. Vehicle recall populations may inform an exposure inventory, especially when joined to a version-level insured roster. They cannot replace crash causation, bodily-injury severity or coverage adjudication data.
The 2025-amended NHTSA crash-reporting order requires specified ADS and Level 2 ADAS crash reports from named entities under different criteria. NHTSA itself warns of nonrepresentative reporting due to unequal telemetry, awareness and duplicate-report issues. These reports have important safety surveillance uses. They do not provide a representative numerator of cyber-induced accidents or a homogeneous exposure denominator. Reporting under the order should never be coded as proof of hostile interference simply because software or automation was active.
Silence, cyber exclusions, and multi-line overlap
Cyber insurance and motor coverage may overlap in causal origin while covering different insured interests. A cyber policy might respond to forensic expenses or digital interruption, while a motor liability policy responds to a road user's bodily injury, subject to respective terms. Some product liability or commercial general liability placements may respond to a technology vendor's legal responsibility. There is no single general rule that physical harm caused through digital systems must belong to a cyber policy. The policy wording and governing law control the practical boundary.
In the 2022 supervisory statement on non-affirmative cyber exposures, EIOPA urged attention to cyber underwriting exposure potentially embedded in insurance lines that do not affirmatively sell such coverage. Motor-related cyber-physical losses are a natural exposure-review question in that setting, but the statement should not be read as a finding that every motor policy provides or excludes cybersecurity loss. For accumulation analysis, a reviewer should identify explicit cyber clauses, silent or uncertain triggers, statutory victim rights, terrorism or war provisions where relevant, ensuing loss conditions, occurrence language, deductibles and policy priority rules.
A useful coverage register records at least the issuing insurer and jurisdiction, insured interest, motor or other policy class, insured perils, relevant cyber wording, statutory override, claimant and beneficiary, trigger date, per-claim and per-event limits, aggregate limit, deductible, defense costs treatment, and possible recovery counterparties. This is an analytic inventory rather than a substitute for legal advice. It enables a catastrophe scenario to flow through each actual contractual layer rather than assigning all cyber-originated physical injury to a generic combined insurance pool.
Reserving, prudential risk, and cross-line financial effects
Technical incident counts, probable legal claims, reportable claims, established compensation duties, case reserves, incurred-but-not-reported claims and paid losses differ in measurement. A common vulnerability can prompt delayed reporting or precautionary inspection without injury, while severe physical injury can produce long-term medical care and legal expenses. An insurer should maintain separate estimates for gross ultimate liability, defense expenses, case and IBNR components, expected recourse, and reinsurance recovery. Cyber event labels can assist segmentation but do not replace claim-level estimation.
The IFRS 17 Insurance Contracts overview explains measurement of insurance contracts and future cash flows under an accounting standard effective for annual reporting periods beginning on or after January 1, 2023. This accounting framework is not itself a cyber accumulation model. Its relevance lies in the requirement that uncertain future obligations be measured through an appropriate contractual and cash-flow lens; detailed measurement judgments require the full standard and relevant accounting policy.
The IAIS Insurance Core Principles include ICP 16 on enterprise risk management for solvency purposes, emphasizing integrated identification, measurement and management of insurer risks and their interdependencies. A motor cyber-physical accumulation review therefore has to connect underwriting data with technology concentration, portfolio overlap, treaty recoverability, model error and potential liquidity strain. Capital analysis that assumes widespread supplier uniformity away through vehicle-level diversification can mischaracterize tail risk, even if average motor frequency is well estimated.
Evidence, Alternative Models, and Methodological Limits
What the primary materials establish
The strongest available evidence is not a single unified estimate of insured motor cyber-catastrophe loss. It is a set of different types of records, each directly supporting a bounded proposition. The 2010 and 2011 original automotive studies establish technical means of manipulating studied systems; the 2015 researcher demonstration and federal recall establish the reality of a remotely reachable vulnerability and a substantial potential installed base; UNECE and NHTSA materials establish security and remediation governance; statutes establish eligible legal pathways; EIOPA and IAIS describe underwriting and prudential concerns. None of those records, alone or in combination, identifies an observed annual loss frequency, a validated malicious-accident attribution rate, or a contemporary insurer's common-shock tail distribution.
A source-to-claim audit must distinguish four evidence classes. Official legislation and regulatory instruments establish the text of rules, subject to legal version and commencement checks. Manufacturer disclosures, recall notices and official safety findings establish documented operational or potential safety conditions, not universal accident outcomes. Peer-reviewed or original experimental security papers establish what happened in specified tests and the tested assumptions. Insurer claims, traffic records and reproducible statistical studies are needed to estimate real-world frequencies, severities and dependencies. The synthetic actuarial calculations presented in this monograph are a fifth object: internally testable implications of assigned assumptions rather than empirical evidence. A valid scholarly argument may connect all of these classes but cannot silently upgrade one into another.
The 2015 NHTSA recall notice is especially instructive. Its count is a potential population of affected vehicles. Multiplying 1.4 million by a hypothetical claim severity and calling the result a probable loss would conflate the stock of vulnerable objects with the frequency of physical accidents. Conversely, dismissing systemic insurer exposure because the notice proves no large crash count would ignore the demonstrated dependency. The warranted conclusion is more precise: a widespread technical weakness can create a concentration susceptible to correlated loss, while its expected insurance consequences require additional probabilities, accident conversions and contract-level evidence.
Denominators, selection, and missing data
A cyber crash rate has meaning only with an appropriate risk-time or distance denominator. Candidate denominators include connected vehicle-years in a particular software configuration, miles driven with a vulnerable feature enabled, exposed trips within a designated automation domain, or operating hours during a compromised infrastructure interval. These are not interchangeable. An insurer may observe vehicle-years but lack firmware and activation histories; a manufacturer may observe firmware deployment but lack detailed insurer claim amounts. The population used in an estimated numerator must be compatible with the denominator in software version, calendar time, country, operational mode and reporting threshold.
NHTSA's Standing General Order data limitations explicitly caution that reporting varies with manufacturer knowledge, telemetry and classification and that some incidents may be duplicated. Automated-driving crash reports are therefore a selected surveillance dataset, not an independent sample of all motor operation and not an indicator that malicious cyber interference caused the reported crash. Even a perfectly deduplicated accident series would still require cyber attribution and matching exposure units. A credible model must specify missingness mechanisms and refuse a causal cyber label when the event record provides only system engagement.
In public cyber-risk datasets, Biener and colleagues' empirical analysis likewise illustrates selection and incomplete information, though its 994 cases are general operational cyber losses, not an automotive bodily-injury sample. Claim-data availability may be particularly limited for settled liability disputes, confidential supplier terms, or events classified as ordinary crashes rather than software-induced accidents. Estimation should therefore use clear uncertainty intervals, track the provenance of labels and report sensitivity to plausible unobserved cyber causes, rather than asserting population representativeness.
Identifiability and alternative explanations
A model of common accumulation requires identification of at least the susceptible fleet, the probability of a common initiating event, the conditional proportion physically affected, the chance that effects convert into compensable accidents, and the distribution of final insured payments. In some data these components are not separately identifiable. If only total paid claims are observed, an increase may reflect more accidents, more claimants per accident, higher injury severity, inflation, policy-limit changes, greater litigation activity, or more vulnerable vehicles. More software and network telemetry cannot resolve every liability or settlement question; more claims data cannot by themselves identify the upstream technical event.
Three alternative explanations warrant explicit competition in incident attribution. First, a common software release may correlate vehicle faults but no external adversary is involved. Second, common traffic weather or driver responses may cluster accidents across vehicles that do not share a digital weakness. Third, the apparent common factor may be a reporting campaign: after a recall, claimants, repairers and investigators may identify and report similar complaints more often. A credible causal investigation compares these mechanisms using chronology, version, geography, traffic and counterfactual information. The mere co-occurrence of software similarity and accident reports is weaker than a demonstrated technical pathway with matched event logs and relevant failure behavior.
The legal component has its own identifiability limits. Two claims arising from identical technical conditions may differ because one vehicle was operating within an authorized automated feature and another required active driver supervision. Two policies may cover different defendants or injury classes. A reinsurance treaty may attach only after an occurrence aggregation decision. A pooled claims table omitting these contract and jurisdiction fields cannot uniquely determine either gross or net insurer liability.
Robustness matrix for accumulation models
The baseline stress calculation should be repeated under explicit perturbations, with each result tied to the changed parameter rather than presented as an unqualified loss prediction.
| Dimension | Baseline assumption | Alternative sensitivity | Consequence to measure |
|---|---|---|---|
| Common event | Annual Bernoulli disturbance probability 1% | Disturbance probability 0.2% or 2% | Frequency of major modeled shock |
| Conditional physical accident | 2% per exposed vehicle | 1% or 10%, with revised event probability | Event size and portfolio tail |
| Exposure | 10,000 equal vehicles | Vulnerable cohort split by software version | Concentration and mitigation benefit |
| Severity | $250,000 fixed gross insured payment | Injury and property mixture with heavy right tail | Occurrence exhaustion and reserves |
| Conditional dependence | Independent accidents given disturbance | Traffic-linked clustering or shared crash chains | Additional covariance and tail concentration |
| Geographic reach | Entire modeled fleet susceptible | Bounded operational domain or regional outage | Exposed fraction and event duration |
| Liability | Every modeled accident results in one covered payment | Disputed causation, limits and different insured parties | Paid-versus-physical loss conversion |
| Reinsurance | All eligible loss grouped into one occurrence | Each accident distinct; competing aggregation clauses | Recoverability and net capital |
| Data | Complete version and loss attribution | Missing telemetry and selective reports | Statistical identification and uncertainty |
Under each sensitivity, analysts should compute the same set of outcomes: expected gross loss, variance, a declared tail quantile or exceedance probability, gross layer payments, possible recoveries under each contract interpretation, and retained loss after recovery collectibility assumptions. Only the first two moments of the simple mixture model have been calculated here; no unstated tail quantile, actual deductible distribution, treaty pricing result or market capital number is claimed.
The baseline and alternatives also identify two separate kinds of uncertainty. Within a specified mathematical model, ordinary random variation can be calculated. Uncertainty about vulnerability prevalence, attacker capability, accident conversion, legal interpretation, and tail claim severity changes the model itself. A very precise conditional simulation of an unvalidated scenario is not a precise prediction. Confidence intervals arising from a fitted independent-accident count model will miss omitted common shocks even when the software implementation is flawless.
Reproducibility and numerical audit
For the simple common-event construction, exact binomial conditioning provides the checkable probability law. With N vehicles, Z distributed as Bernoulli(q), Y_i as independent Bernoulli(p), and a constant severity s, the claim-count probability at zero is P(C = 0) = (1−q) + q(1−p)^N. For integers c from 1 to N, P(C = c) = q times the binomial probability for N trials and p. This two-part law explains why zero annual losses can remain likely while a rare realized shock generates a large number of claims. Under an independent marginal approximation with event probability qp, the zero probability would instead be (1−qp)^N, which is generally different.
The law of total variance yields Var(C) = E[Var(C|Z)] + Var(E[C|Z]) = qNp(1−p) + q(1−q)(Np)^2. The corresponding aggregate insured payment has E[G] = sE[C] and Var(G) = s²Var(C). These formulas can be recomputed by direct finite binomial summation for any realistic N using stable probability algorithms or verified against an appropriately seeded simulation. The independent marginal comparator has Var(C_ind) = Nqp(1−qp). For the baseline, the difference in count variance equals N(N−1)q(1−q)p², exactly the contribution of cross-vehicle covariance summed over distinct pairs.
These identities establish internal mathematical consistency of the synthetic exercise; they do not verify q, p or s in any observed fleet. Numerical validation should maintain a ledger of assumptions, exact arithmetic, software environment and failures and should not insert point estimates of genuine insurer loss without authoritative evidence. The reproducibility result is the conditional algebra and a bounded illustrative sensitivity experiment, not a new predictive method.
Source-access and legal-version limits
The UNECE regulation landing pages and ISO catalogue establish instrument identity and regulatory or engineering scope. They do not substantiate claims that every jurisdiction has enacted identical rules or that any specific manufacturer passed a conformity audit. The public IFRS overview confirms the standard's nature and broad accounting principles but does not replace the full licensed accounting standard. Similarly, the official UK legislation confirms general rules with section-specific commencement provisions that must be verified at the accident date. Specific American state tort holdings and motor policy endorsements are outside this dossier's primary-source comparison; no national U.S. coverage holding is therefore stated.
The corrected EU 2024 Product Liability Directive text is particularly time sensitive: the article's corrected 'after December 8, 2026' product-entry date is different from the originally printed 'after December 9, 2026' text. The prospective date is stated as a law-version fact, not as an assertion that national measures had already been implemented on November 16, 2026. Future claims research must attach a specific national transposition and cause-of-action date before making liability determinations.
Comparative Scientific Synthesis and Research Conclusions
What integrates across countries and lines of insurance
Several conclusions survive differences in country, vehicle category and contractual regime. A technical vulnerability can be widely shared without having caused any proven crashes. A motor crash can generate an insured victim-compensation duty even when ultimate technical responsibility is disputed. A manufacturer's exposure is not necessarily the same as the motor insurer's gross payment. A common supplier or update service can connect an otherwise diversified set of insured vehicles. And the gross payment, reinsurance recovery, reserve estimate and final net economic burden must be computed at different stages with distinct evidence. These are analytically defensible results of the source comparison, not proclamations of a new system of insurance theory.
The most consequential insurer-level result is the inadequacy of vehicle independence when a common initiating event affects substantial portions of a portfolio. The illustrative model does not predict a real cyber catastrophe; it proves a conditional proposition: under a shared disturbance and otherwise identical marginal annual claim probability, aggregate variance exceeds its independent-risk counterpart by the sum of positive pairwise covariances. The expected payment remains unchanged. This distinction explains why a model fitted successfully to average motor loss may still produce materially weak tail or reinsurance estimates when software dependency is omitted.
The comparison of EU, UK and U.S. legal material supports a second conclusion. A single technical cause can traverse different liability channels. The EU emphasizes compulsory vehicle-use liability protection with complementary product-defect rules; Great Britain has enacted a specified insurer-first route for automated driving with detailed conditions and a newer authorization framework; the United States retains material state variation in motor insurance and private civil liability alongside federal vehicle safety oversight. These are different legal entry points into compensation, while the physical underlying injury may be the same. No universal cross-border coverage rule can be inferred from their coexistence.
An important practical implication concerns transaction-level financial records. Gross claimant payments, legal settlements, contribution receipts, reinsurance recoveries and other vendor indemnities must be linked to a common underlying harm so that portfolio and market aggregates do not double count. In any insurer, a claim outcome register should also preserve uncertain defense and recovery positions. Financial aggregation should answer at least four separate questions: what physical harm occurred; who legally owes compensation; what amount each insurer must initially fund; and how that cost is subsequently distributed among insurers, reinsurers and responsible third parties.
How mitigation changes the insurance distribution
Mitigation is not a scalar discount applied mechanically to every scenario. Strong update controls may reduce the vulnerable installed base; segmentation may block the path from infotainment to control systems; anomaly detection may shorten a compromise; safe fallback can prevent a control error from becoming a collision; rapid recall can reduce exposure days; better logs can increase confidence in causation and recovery. These interventions modify different parameters and dependencies in the loss chain. A single measure can reduce the probability of a common shock, the conditional number of affected vehicles, the accident conversion probability, or claims uncertainty, and its contribution should be measured at the stage it actually changes.
The relevant standards and public guidance point in these directions. UN R155's cybersecurity management framework and NHTSA's 2022 safety guidance address organizational and technical risk control, while UN R156's software update framework focuses on an important path of changing exposure. For insurers, engineering control assurance should be recorded with its test date, vehicle population and technical boundaries. Inferring that standards compliance eliminates accumulation risk would exceed what these materials establish.
Diversification must also be redefined operationally for the cyber-physical context. Insured vehicles geographically dispersed across many roads may still use one vulnerable software module, cloud service, vehicle platform or signing authority. Conversely, vehicles of the same manufacturer but on independently maintained and segmented systems may have less common-mode exposure than an undifferentiated manufacturer count suggests. An exposure inventory organized only by make and model is therefore an imperfect proxy for the dependency structure that drives catastrophe accumulation.
Evidence priorities for independently testable work
The first priority is a versioned, privacy-conscious insured fleet inventory connecting vehicle identity to hardware, software, automation features, service provider, geography, operating time and policy coverage. The second is a causally adjudicated accident archive with technical evidence, human-control context, alternative-cause assessment and quantified bodily injury and property damage. The third is a financial ledger linking accident, claimant, policy, reserve, payment, expense, subrogation, treaty and ultimate net recovery. The fourth is a documented system-dependency map covering vendors, common services, credential arrangements, update channels and infrastructure, with confidence ratings and effective dates. These are proposed research data requirements, not a claim that an existing proprietary data source already fulfills them.
Validation can proceed by comparing historical nonmalicious software defect cohorts, matched recall populations, controlled engineering tests and specifically substantiated cyber-origin incidents. The datasets should remain distinguishable because their causes, selection processes and loss outcomes are different. Counterfactual questions include the accident rate of otherwise similar vehicles without the implicated software version, the effect of a completed patch on actual exposure, and the fraction of hazardous control effects prevented by safe-fallback features. A trustworthy design requires sufficiently detailed records and explicit acknowledgment of unobserved conditions; a recall alone is insufficient to estimate these effects.
The model-comparison program should assess whether independent Bernoulli claims, latent common-shock mixtures, random-effect frequency models, dependency-aware event simulations and contractual event aggregation produce materially different capital-relevant outputs under the same evidence. Model selection should be made by calibration, out-of-sample validation, stress behavior, transparent assumptions and interpretability of insured cash flows. Introducing a new branded actuarial theory would not improve the evidence base. The valuable research task is to make existing probabilistic and legal machinery answer the correct joint technical-financial question.
Final analytical findings
The verified technical record demonstrates attack and fault pathways that can influence vehicle safety; the official 2015 recall illustrates how a single software-related vulnerability can span a large installed population without documenting equivalent mass injury. The legal record demonstrates that the first payer, ultimate responsible party and insured contract can differ. The mathematical record demonstrates that common-source dependence changes the dispersion and tail structure of insured losses even with an unchanged average claim cost. The supervisory record supports identifying, measuring and managing embedded or silent cyber exposures and aggregating correlated risks at enterprise and group level.
These findings together support a practical scientific conclusion: credible vehicle cyber-physical motor liability accumulation analysis requires a causally traceable bridge among shared technical dependency, physical incident, jurisdiction-specific compensation, policy-level cash flows and reinsurance interpretation. The bridge is assembled from verified records and tested assumptions. Its quality is measured by how precisely each link can be evaluated and how explicitly unresolved evidence affects the range of possible outcomes.
Conclusion
Vehicle cyber-physical failure is a potentially correlated motor liability peril because common software and infrastructure dependencies can transmit one initiating condition to many otherwise separate vehicles and insurance contracts. The probability of technical exposure, accident conversion, compensable injury, covered liability, and reinsurance recovery are separate variables. The engineering literature and official safety records establish plausible and in some cases demonstrated technical pathways; they do not establish an observed global motor cyber-catastrophe loss distribution. Comparative statutes and insurance supervisory sources establish the need to identify victim-compensation mechanisms, product claims, first-instance insurer duties, contract limits, recoveries and their effective legal dates.
The transparent 10,000-vehicle calculation gives a precise conditional demonstration: a portfolio with $500,000 expected annual gross loss can have a common-event conditional expected loss of $50 million and greatly elevated variance compared with a model that assumes independent accidents at the same marginal frequency. Altering the shock probability and severity while holding their product fixed changes conditional event size and capital-sensitive exposure. These figures are explicitly synthetic; their scientific value lies in revealing the assumptions that must be measured in real fleets, not in predicting present insurance claims.
The complete source comparison establishes the research priorities of version-level exposure records, incident-level causal evidence, jurisdiction-specific insurance interpretation, measured cross-portfolio dependencies and contract-specific reinsurance aggregation. An insurer that integrates those layers can investigate physical risk and financial accumulation using auditable facts and explicit stress assumptions. The monograph supplies an original comparative and analytical synthesis of established results, with no new named theory, doctrine or estimator and no assertion of externally peer-reviewed novelty.
References
Biener, C., Eling, M., & Wirfs, J. H. (2015). Insurability of cyber risk: An empirical analysis. The Geneva Papers on Risk and Insurance – Issues and Practice, 40(1), 131–158. DOI: 10.1057/gpp.2014.19 Original source
Böhme, R., & Kataria, G. (2006). Models and measures for correlation in cyber-insurance. Fifth Workshop on the Economics of Information Security (working paper, revision 0.3). Original source
Checkoway, S., McCoy, D., Kantor, B., Anderson, D., Shacham, H., Savage, S., Koscher, K., Czeskis, A., Roesner, F., & Kohno, T. (2011). Comprehensive experimental analyses of automotive attack surfaces. Proceedings of the 20th USENIX Security Symposium. Original source
European Insurance and Occupational Pensions Authority. (2022). Supervisory statement on the management of non-affirmative cyber exposures. EIOPA. Original source
European Parliament & Council. (2009, amended through 2023). Directive 2009/103/EC relating to insurance against civil liability in respect of the use of motor vehicles (consolidated version, December 23, 2023). Original source
European Parliament & Council. (2019). Regulation (EU) 2019/2144 on type-approval requirements for general safety of motor vehicles, as amended. Original source
European Parliament & Council. (2021). Directive (EU) 2021/2118 amending Directive 2009/103/EC on motor vehicle liability insurance. Original source
European Parliament & Council. (2024, corrected 2026). Directive (EU) 2024/2853 on liability for defective products (consolidated text with 2026 corrigendum). Original source
European Union Agency for Cybersecurity & European Commission Joint Research Centre. (2021). Cybersecurity challenges in the uptake of artificial intelligence in autonomous driving. ENISA. Original source
IFRS Foundation. (2017, effective 2023). IFRS 17 Insurance Contracts. Original source
International Association of Insurance Supervisors. (2024). Insurance Core Principles and ComFrame, adopted revisions December 2024 (online tool). Original source
International Organization for Standardization & SAE International. (2021). ISO/SAE 21434:2021, Road vehicles—Cybersecurity engineering. Original source
Koscher, K., Czeskis, A., Roesner, F., Patel, S., Kohno, T., Checkoway, S., McCoy, D., Kantor, B., Anderson, D., Shacham, H., & Savage, S. (2010). Experimental security analysis of a modern automobile. 2010 IEEE Symposium on Security and Privacy, 447–462. DOI: 10.1109/SP.2010.34 Original source
Miller, C., & Valasek, C. (2015). Remote exploitation of an unaltered passenger vehicle (white paper). IOActive. Original source
National Association of Insurance Commissioners. (n.d.). Auto insurance: Consumer information. Original source
National Highway Traffic Safety Administration. (2015). Safety recall 15V-461: Radio software security vulnerabilities (recall acknowledgement, July 24). U.S. Department of Transportation. Original source
National Highway Traffic Safety Administration. (2022). NHTSA updates cybersecurity best practices for new vehicles (September 7). Original source
National Highway Traffic Safety Administration. (2025, updated data through 2026). Standing General Order on crash reporting: ADS and Level 2 ADAS. Original source
United Kingdom. (2018). Automated and Electric Vehicles Act 2018, Part 1, section 2 (as amended). legislation.gov.uk. Original source
United Kingdom. (2024). Automated Vehicles Act 2024 (as amended; commencement to be checked section by section). legislation.gov.uk. Original source
United Nations Economic Commission for Europe. (2021). UN Regulation No. 155—Cyber security and cyber security management system. Original source
United Nations Economic Commission for Europe. (2021). UN Regulation No. 156—Software update and software update management system. Original source
World Health Organization. (2023). Global status report on road safety 2023. WHO. Original source