Emmanuèle Lutfalla and Louis Fer discuss the emerging insurance risks posed by artificial intelligence, in Cyber Security Intelligence

The cybersecurity industry learned the hard way what supply chain concentration means. SolarWinds compromised eighteen thousand organisations through a single software update. Log4Shell exposed virtually every enterprise running Java. MOVEit turned a file transfer tool into a mass breach vector.

Each time, the pattern was the same: a shared dependency failed, and thousands of unrelated businesses discovered simultaneously that they were connected by a vulnerability none of them had individually assessed.

Artificial intelligence is creating the conditions for the next iteration of this problem but at a scale and across a range of business functions that prior supply chain incidents did not reach.

The Model Layer as Critical Infrastructure

Enterprise spending on frontier AI grew by over 300 per cent in 2025. What that figure obscures is that most of it flows to a handful of providers. A small number of foundation models now sit inside the operations of banks, healthcare providers, law firms, manufacturers and insurers themselves. Unlike a conventional software dependency, these models do not simply process data. They generate outputs, make assessments and, increasingly, initiate actions, drafting legal opinions, approving credit decisions, triaging patient referrals.

From a cybersecurity standpoint, this creates a threat surface that existing risk frameworks struggle to categorise. An adversarial attack on a widely deployed model's API infrastructure, a poisoned training dataset propagated through a routine update, or a coordinated prompt injection campaign targeting a foundation model's reasoning layer could produce cascading failures across sectors and jurisdictions simultaneously. The difference from SolarWinds is not merely one of scale. It is that the compromised system is making decisions, not just transmitting data.

Where Cyber Insurance Stops & the Gap Begins

Cyber insurance was built around a recognisable anatomy: unauthorised access, data exfiltration, ransomware, business interruption from IT downtime. The wordings reflect this. When an AI model produces a flawed credit assessment that costs a bank's client several million pounds, is that a cyber event? When an AI-powered diagnostic tool misclassifies a condition and the hospital faces a malpractice claim, which policy responds: cyber, professional indemnity, or product liability?

The honest answer is that no one is certain, and current policy language was never designed to resolve the question. The industry has begun describing this as "silent AI", unintended exposure embedded in conventional wordings that neither expressly cover nor expressly exclude AI-driven losses. Research co-authored with Anthropic and OpenAI earlier this year estimated that more than 90 per cent of insurers' total AI liability exposure sits in this uncharted territory.

The silent cyber experience is instructive but insufficient as a precedent. When the market recognised that property and business interruption policies were silently absorbing cyber risk, it introduced exclusions, clarified wordings and developed standalone products. That effort took years and is arguably still incomplete. Silent AI is a harder problem because it does not map to a single product line. It touches professional indemnity, general liability, D&O, technology E&O, and potentially every casualty class on a carrier's books.

The Causation Puzzle

For security professionals accustomed to incident response, AI failures raise a classification problem that ripples directly into insurance coverage. Consider a scenario in which a widely used AI model is compromised through a supply chain attack. The model begins generating subtly incorrect outputs across its user base. A law firm relies on those outputs in advising a client. The client suffers a financial loss.

The insurer must determine: was the proximate cause the security breach? The firm's failure to supervise the AI? The model provider's inadequate security controls? The client's reliance on advice they knew involved AI?

The answer determines which policy, if any, responds, and potentially which exclusion applies. Multiply that scenario across hundreds of insureds sharing the same compromised model, and the scale of the coverage dispute becomes apparent.

European courts will face these questions with limited legislative guidance. The AI Liability Directive, which was intended to harmonise civil liability rules for AI across the EU, was withdrawn by the Commission in 2025. National judges are left to apply traditional fault-based doctrines to autonomous systems, doctrines developed for human conduct and ill-suited to algorithmic decision-making. In the United States, ISO introduced generative AI exclusion endorsements for commercial general liability in January 2026, but adoption has been fragmented. Large parts of the market remain unaddressed.

What the Security Community Should be Paying Attention To

CISOs and risk managers tend to evaluate AI through the lens of data protection, model governance and adversarial robustness. These are necessary concerns, but they miss the insurance dimension, and it is the insurance dimension that will determine who absorbs the financial impact when an AI system fails at scale.

The challenge facing the insurance market is that the tools available for assessing and transferring AI risk remain at an early stage of development. Few carriers have yet implemented systematic processes for identifying which AI models their insureds depend on, or for mapping common dependencies across a portfolio. Reliable methods for estimating correlated AI losses are still being developed. When the first large-scale AI incident occurs, whether through an adversarial attack, a defective update, or a failure mode no one anticipated, the resulting claims will be fought over wordings that were drafted for a different era of technology risk.

The precedent from terrorism is sobering. After September 11 generated approximately $47 billion in insured losses, the private market did not innovate its way through the problem. It withdrew capacity, and governments were compelled to step in through mechanisms like TRIA in the United States and GAREAT in France. Whether the insurance market can develop adequate capacity for AI tail risk, or whether it will again need to withdraw and recalibrate, depends in large part on what happens next.

The minimum requirement is transparency, both from AI providers about their security architecture and update practices, and from insureds about their AI dependencies. Without that information, neither the cybersecurity community nor the insurance market can credibly assess the exposure they are carrying.