Insurers Filed Deepfakes Under Cybersecurity. The Risk Walked Into The Claims File.

Estimated reading time: 10 minutes

It is a Tuesday morning in early 2027. At 9:04, a policyholder’s AI agent files a first notice of loss with your carrier. Forty seconds later, your own AI agent has triaged the claim, requested supporting evidence, summarised the file, and drafted a settlement recommendation. Fast. Frictionless. Fully auditable. And nobody has asked the question the business lives or dies on.

The water damage photograph was doctored. It took twelve seconds to make, using a model that cost the policyholder nothing.

That scenario opens The Speed of Trust, published on 3 September by insurance innovation strategist Sabine VanderLinden of Alchemy Crew Ventures and commissioned by voice-based risk assessment firm Clearspeed. The Tuesday morning is imagined. What the researchers found in the filings is not.

They searched 76 annual reports, 10-K filings, proxy statements, and statutory returns from 49 insurers and reinsurers for a dozen terms relating to AI-generated and manipulated evidence.

Synthetic media returned nothing. Synthetic identity returned nothing. Voice cloning returned nothing.

Deepfakes did marginally better. Six of the 49 companies mentioned them. All six filed the risk under cybersecurity. None connected it to the evidence used in claims or underwriting decisions. Five of the world’s top ten reinsurers sat among those analyzed. None mentioned deepfakes, synthetic media or AI-generated evidence in their most recent annual reporting.

Alongside the filings review, the research draws on 31 industry studies and 16 interviews with claims and underwriting leaders across the United States and United Kingdom.

The Verification Gap Behind Deepfake Insurance Claims

The report’s central number comes from Verisk research published in March 2026, based on a survey of 300 US claims professionals. Ninety-eight percent agree AI editing tools are driving a rise in digital media fraud. Thirty-two percent are very confident they could identify a deepfake.

VanderLinden calls the space between those two figures the verification gap. “Insurance is automating decisions faster than it can verify the information behind them,” she said.

The cultural picture underneath it is shifting too. Verisk found that 55% of Gen Z and 49% of Millennials would consider a small edit to a claim photo, compared to 12% of Baby Boomers. Sixty-two percent of consumers now believe document manipulation is common.

The threats themselves are not unknown to the industry. Munich Re’s 2026 outlook named deepfakes, voice clones and synthetic identities as supporting fraud and social engineering at scale. The World Economic Forum found cyber-enabled fraud overtaking ransomware as the threat CEOs worry about most. The filings analysis shows that this recognition has not carried over from outlook reports into the documents where carriers set reserves.

Why Better Automation Produces Worse Trust

The report frames the problem as a paradox rather than a technology failure.

Agentic AI executes tasks correctly. It cannot assess whether the task should be trusted. A faster, cleaner workflow executes flawlessly on a bad input, and removes the human friction that used to surface it. McKinsey research cited in the report found only about a third of organizations have mature controls for the agentic era.

Scott Clayton, claims fraud leader at Zurich UK, gives the practical version. Someone can photograph a person at dinner, ask a consumer AI tool to add a Rolex to their wrist, and the tool will simply do it.

That corresponds to what CyberCube set out on 1 September from the modeling side, mapping six families of AI-driven loss across eleven policy types and declining to say which policies respond. Two reports, days apart, the same gap approached from opposite directions. It also sits alongside red-team research on autonomous agents showing agents taking consequential actions that nobody authorized.

Incident response data tells a slower story on pace. Sophos found no fully autonomous AI-driven attacks across 661 cases in its 2025 casework and confirmed a single deepfake incident, reported quickly and contained. Its conclusion was that AI is adding scale and noise rather than replacing attackers. The two views are not incompatible. One measures what has already reached the claims file. The other measures what carriers are prepared for when it does.

A Court Has Already Run The Real Version

The report’s Tuesday morning is hypothetical. A California courtroom has already handled the actual thing.

In Mendones v. Cushman & Wakefield, self-represented plaintiffs in a housing dispute moved for summary judgment and supported it with video they said showed a witness testifying. Judge Victoria Kolakowski noticed the expression barely moved. The voice did not track the words. The same gestures repeated.

Visual inspection is not what broke it. The video’s metadata identified an iPhone 6 as the recording device, contradicting the plaintiffs’ account of how they captured the footage. The court found the evidence fabricated and dismissed the case with prejudice.

Cyber Insurance News examined that ruling in August alongside two other cases. This case comes closer than any on record to an insurance claim involving fabricated supporting media, and the file carried an unintended detail that exposed the deception.

Mike Nelson, vice president and field chief technology officer at DigiCert, framed the insurance relevance in that reporting. “Cyber claims reports depend on reconstructing a timeline: who initiated a transaction, what communication was received, when systems became unavailable, and what changed during an incident,” he said. He named business email compromise, deepfake-enabled impersonation, fraudulent payment instructions, ransomware and business interruption as the places verifiable provenance would matter most.

On whether underwriters will start demanding it, Nelson was deliberately measured. “This will likely be risk-based, not universal,” he said, with attention to who can create and sign content and whether provenance survives editing and distribution.

Worth noting for anyone waiting on regulation: both Mendones and the earlier Washington ruling in State v. Puloka were decided under existing evidence rules. Federal Rule of Evidence 707 and the companion deepfake amendment to Rule 901 remain proposals. Judges have not been waiting.

Nelson will appear on an upcoming episode of the Cyber Insurance News podcast.

Take The Cyber Insurance News Survey

The Cost Of Treating Everyone As A Suspect

The report’s more provocative argument is that fraud detection has become the wrong instrument.

Legacy fraud models flag between 30% and 50% of claims. Special investigation units have the capacity to work 2% to 5%. The report calls it a backlog generator rather than a detection system, and argues that the people paying for it are the genuine majority waiting behind flags that nobody can work.

Ian Thompson, Zurich Insurance’s former group chief claims officer, told researchers that upwards of 90% of claimants are honest. They simply need insurers to handle their claims quickly. Thompson asks how many believe insurers distrust them because the claims process targets the other 10%.

The Coalition Against Insurance Fraud puts fraud at roughly 10% of property and casualty losses and estimates that $308.6bn is drained annually from the US system. The report’s position is that insurers pay for trust twice: once through leakage and again through friction imposed on everyone else.

A $25,000 Claim And A $7m Verdict

Dale Diamond, vice president of claims at NAMICO, gave researchers a case that shows what happens when nobody looks twice.

A routine trucking claim entered the system at $25,000. It left as a $7m jury verdict, 280 times what it started at.

The instructive part is not the ratio. It is the sequence. The file passed four separate review points. The first rested on a single adjuster’s opinion and was never roundtabled. A brain injury indicator surfaced in the medical records, and nobody escalated it. Liability was never in doubt, and there was no mediation. The case met a jury before anyone had tested how it would land.

Nobody was careless. As Diamond told researchers, a claim rarely goes bad because of one thing. It goes bad through a series of unfortunate events.

The report’s argument is that each of those four moments was a window, and that a signal flagging which interactions deserve a second set of eyes is cheapest while the window is still open.

What A Trust Intelligence Layer Would Do

VanderLinden’s proposal is to make trust a measurable operating layer rather than a late-stage judgment.

The Trust Intelligence Layer, as the report describes it, is a continuous risk indicator running across the policyholder journey, from application and bind through renewal, endorsement, first notice of loss and settlement. Its purpose is to clear the genuine majority quickly and route human judgment to the exceptions.

Three design constraints matter for anyone assessing it. The signal informs a decision rather than making one. It produces an audit trail rather than an automated denial. And it requires no demographic or historical knowledge of the person being assessed.

Clearspeed’s implementation applies signal processing to vocal responses. Respondents answer a short set of automated questions in any language, with results delivered instantly. The report is explicit that this is not voice-stress analysis, not biometric voiceprinting, and not emotion recognition.

The Underwriting Case, And Where Reinsurers Feel It

The report argues the bigger opportunity sits earlier than claims.

Renewal books drift. A household adds a teenage driver. A vehicle moves address. A property’s occupancy changes. A business outgrows the exposure it was priced for. Carriers leave profitable renewal books alone while hidden risk accumulates underneath. Of the filings reviewed, only three mentioned premium audit at all, and none strategically.

Comparison sites and embedded distribution remove the human conversation where a follow-up question would once have surfaced the change.

The report’s reinsurance argument runs through reserve development. It cites Everest Group’s 2024 10-K disclosure of $684m in unfavorable prior-year casualty reserve development in its reinsurance segment, with a further $1.1bn in insurance, attributed to social inflation and US casualty exposure. Cleaner facts at bind, the argument goes, result in fewer surprises inside the treaty later.

On the numbers behind that case, the report is careful with its own weakest data. Actuarial modeling of the US homeowners book estimating 3.7 to 4.2 combined ratio points of leakage is flagged twice as a Clearspeed-affiliated analysis and described as directional rather than independent proof.

Vision 2030

The report closes on agent-to-agent transactions.

By 2030 it expects a material share of insurance interactions to run between a customer’s AI agent and a carrier’s AI agent, at machine speed, with no human in routine business. The verification question does not disappear in that world. It migrates. When the action is always executed correctly, what remains is whether the interaction behind it can be trusted.

The report’s phrase for the shift is moving from trusting the human to trusting the conversation. Its closing judgment is that the arms race is symmetrical, and the only durable advantage is establishing trust faster than an adversary can manufacture doubt.

For underwriters and claims leaders, the practical takeaway sits in that filings finding. Six mentions of deepfakes across 49 companies, all of them filed under cybersecurity. The exposure is not being named in the documents where reserves get set.

FAQ – Deepfake Insurance Claims

What did the filings analysis find?

Researchers searched 76 filings from 49 insurers and reinsurers for terms relating to AI-generated evidence. Synthetic media, synthetic identity and voice cloning returned zero mentions. Six companies mentioned deepfakes, all framing it as a cybersecurity concern rather than a claims or underwriting issue.

What is the verification gap?

The distance between recognizing a risk and being able to detect it. Verisk found 98 percent of US claims professionals agree AI editing tools are driving digital media fraud, while only 32 percent are very confident they could identify a deepfake.

Why does automation make the problem worse?

Agentic AI executes tasks correctly but cannot judge whether a task should be trusted. Faster workflows execute flawlessly on bad inputs and remove the human friction that previously surfaced them. McKinsey found roughly a third of organizations have mature controls for the agentic era.

What is a Trust Intelligence Layer?

A continuous risk indicator running across the policyholder journey from application through claim settlement. It is designed to clear genuine customers quickly and route human judgment to exceptions. The signal informs decisions rather than making them and produces an audit trail.

How does this affect underwriting rather than claims?

Renewal books drift as households, vehicles, properties and businesses change without the carrier learning of it. Only three of the 76 filings reviewed mentioned premium audit. The report argues cleaner facts at bind reduce the reserve development reinsurers later absorb.

Related Cyber Insurance Posts

• IGI Backs A Cyber Reinsurance MGA Built To “Lead”, Not Follow
• Top Strategies for Identity Verification in the Age of Deepfakes, Remote Work, and AI Threats(Opens in a new browser tab)
• Local Media Reports on Cyber Insurance Issues Involved in Hack of Oakland(Opens in a new browser tab)
• Retailer M&S to Make $133 million Cyber Insurance Claim: Media Reports(Opens in a new browser tab)
• Cyber Insurers Grapple with AI Deepfakes(Opens in a new browser tab)