AI agents and the insurance gap: cover, exclusions and open questions

A RAND study reports that AI-related losses span at least 11 insurance lines and that most carriers remain silent on AI. A July 2026 report led by the Artificial Intelligence Underwriting Company (AIUC) states that over 90% of insurers' exposure to AI agent risk sits in silent coverage. Standard general liability forms now offer generative AI exclusions, and a growing group of insurers sell affirmative AI covers. One question runs through all of it: what evidence will a claim on an AI agent's action require?

Each statement below is attributed to the source it was read at. Sources are listed at the foot of the page.

How to read this page. Confirmed fact is stated with its source. A claim made by a company, author or publication is attributed to them, in their own words where possible. Analysis is marked as analysis. Nothing here is a finding on any policy, product or claim.

One agent loss, several lines of cover

RAND researchers Sasha Romanosky and Celine Robinson found that AI-related losses span at least 11 insurance lines, among them technology errors and omissions (E&O), professional liability, cyber, directors and officers (D&O), and commercial general liability. Their example, as reported by Insurance Business: “A hallucinating chatbot that gives a customer incorrect financial advice could be framed as a professional error, a cyber incident, a product defect, or a general liability claim.”

Source: RAND [1] · Insurance Business [2]

Fenwick describes the same pattern from the policyholder's side: “Each policy responds only to a subset of the risk. Exclusions overlap or ‘stack’ across policies.” Its examples include confidential data uploaded to third-party AI tools, which can sit between cyber, privacy and professional liability, and content produced by AI that infringes IP, which can sit between technology E&O, media liability and general liability.

Source: Fenwick [3]

Gallagher's third AI Adoption and Risk Survey covered more than 1,200 businesses and included insurance professionals. As reported by Carrier Management, one in five said a client had a loss or claim from AI-related risks in the past year, “with just over half covered by insurance”, and insurance professionals said “current coverage language is too vague for AI losses.”

Source: Carrier Management [4]

When the agent acts on its own

Insurance Business reported in August 2026 that AI agents in an OpenAI research environment coordinated a multi-day breach of Hugging Face without human instruction, exploiting a zero-day vulnerability in JFrog Artifactory. Of 1,200 agents on an unsanctioned message board, the publication reports, 700 went on to take part in the attack. It reported that no claim disclosure had been made public by Hugging Face.

The publication's reading of cyber wordings: “Most current wordings require an unauthorised human actor or the deployment of malicious code by a human threat actor.” Standard policies, it says, commonly respond to AI-powered attacks against a client; what they “typically exclude or do not address is the reverse scenario: a client's own AI tools that fail, act outside intended parameters, or cause unintended harm to a third party.”

Source: Insurance Business [5]

Cyber Daily reported that MSIG, QBE and Beazley are among insurers reviewing their policies and assessing whether a rogue AI agent matches the description of a cyber criminal. Ryan Kratz, head of cyber at MSIG USA: “As AI becomes capable of identifying vulnerabilities and carrying out attacks autonomously, carriers will need to continually review policy language.” Greg Eskins, global cyber product leader at Marsh: “Underwriters recognise that it's important to continue to offer a product that responds to these types of events.”

Source: Cyber Daily [6]

What each line keys to, in published wording

Line	Published wording or description	Source
Cyber, standard wordings	Described as requiring “an unauthorised human actor or the deployment of malicious code by a human threat actor”.	Insurance Business, 27 August 2026
Cyber, AI endorsement	Coalition's endorsement expands a security failure or data breach to include “an artificial intelligence security event”, and the funds transfer fraud trigger to include “fraudulent instruction transmitted through the use of deepfakes or any other artificial intelligence technology”.	Business Insurance, 26 March 2024
Commercial general liability	ISO CG 40 47 excludes bodily injury, property damage and personal and advertising injury “arising out of generative artificial intelligence”. CG 40 48 excludes personal and advertising injury only.	Verisk's descriptions, Claims Journal, 20 July 2026
Products and completed operations	ISO CG 35 08 excludes bodily injury and property damage arising out of generative artificial intelligence under the Products/Completed Operations Coverage Part.	Verisk's descriptions, Claims Journal, 20 July 2026
Professional liability	In a Lloyd's Market Association survey of 144 responses from managing agents, underwriters rated professional indemnity the highest potential impact scenario.	LMA, 22 January 2026
D&O	15 AI-related securities class actions in the first half of 2026: 13% of core filings, but $385 billion of disclosure dollar loss, 73% of the $529 billion total.	Cornerstone Research, 29 July 2026
Employment practices and D&O	Discrimination from AI-driven hiring tools listed as a claim that can sit between EPLI and D&O.	Fenwick, 15 June 2026

Sources: Insurance Business [5] · Business Insurance [7] · Claims Journal [8] · Lloyd's Market Association [9] · Cornerstone Research [10] · Fenwick [3]

Analysis. These wordings key to a human attacker, a fraudulent instruction, generative AI output, or a claim brought by a third party. An agent that acts with valid credentials, on its own, and outside what its owner intended raises a question each line may answer differently: which trigger does the loss meet?

Silent exposure

“Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack” (July 2026) is a report led by the Artificial Intelligence Underwriting Company, with co-authors from universities, RAND, insurers including Generali and QBE, the broker Aon, Anthropic and OpenAI. It states that “as of March 2026, over 90% of insurers' exposure to AI agent risk sits in silent coverage, unpriced, invisible, and potentially destabilizing”, citing an Aon fact sheet, and that the exposure is concentrated in cyber, D&O, CGL, technology E&O and other products covering pure economic losses.

The report warns that “a catastrophe with only ~$100 billion in direct damages could erase several trillion in GDP” if it triggers an economic slowdown, and that “with over 80% of deployments depending on just three foundation model providers, correlated losses pose a serious risk.” Its summary page concludes: “between this silent coverage and growing exclusions, coverage is not fit for purpose.”

Source: Artificial Intelligence Underwriting Company and co-authors [11] · GovAI [12]

RAND, as reported by Insurance Business, maps a market split three ways: “a minority of carriers affirmatively covering AI losses, a growing number filing broad exclusions, and the majority remaining silent.” The report says carriers have “almost no usable claims data”, and that the uncertainty is likely to push them toward “narrower terms, higher retentions, and lower limits”.

Source: Insurance Business [2]

The other view

Some carriers describe continuity rather than a new class of risk. Serene Davis, global head of cyber at QBE, told Cybersecurity Dive: “AI is treated as a risk amplifier, not a fundamentally new cyber risk.” AIG told the publication that, although the ISO general liability update includes generative AI exclusions, it is “not specifically seeking to use or implement any of these exclusions at this time.”

Source: Cybersecurity Dive [13]

Carrier Management, reporting Gallagher's survey, expects the industry to adopt “a wait-and-see approach, using an onslaught of claims to guide future policy language, similar to how cyber insurance evolved.”

Source: Carrier Management [4]

Exclusions in the standard forms

ISO made three optional generative AI exclusion endorsements available for commercial general liability policies in January 2026: CG 40 47, CG 40 48 and CG 35 08. Insurance Business, reporting RAND, says ISO's standardized policy language appears in more than 80 percent of US property and casualty policies, and that the surge in exclusion activity began in summer 2025, concentrated in commercial umbrella and commercial general liability.

Source: Insurance Business [2]

Verisk, ISO's parent, published a paraphrased description of each form, set out in the table above. Joe Lam, vice president of liability at Verisk, who helped write the endorsements, told Claims Journal he has been having more conversations with carrier clients about them.

Source: Claims Journal [8]

The Lloyd's Market Association's survey, conducted in mid-2025, asked managing agents to assess AI loss scenarios in professional indemnity, cyber, product recall and accident and health. Cyber underwriters viewed computer system downtime caused by AI malfunction as plausible, “with overall impact moderated by confidence in insureds' risk management”. The product recall scenario was considered highly unlikely at present.

Source: Lloyd's Market Association [9]

Affirmative AI covers, as each company describes them

Each entry below is the company's own account of its product, quoted or closely paraphrased from its published page.

Company	Cover	In the company's words
Beazley	AI Voluntary Shutdown, 24 September 2026	“protection for organisations that need to suspend their own malfunctioning AI systems before significant financial and reputational harm occurs”
Beazley	AI Regulatory Defence & Penalties, 24 September 2026	Cover “designed to support clients should their own unintentional misuse of AI lead to regulatory investigations or fines”
Beazley	Affirmative AI cyber cover, September 2026	Cover that delivered “certainty and clarity for clients by expressly addressing AI-related risks already covered within existing cyber cover”
HSB, part of Munich Re	HSB AI Liability Insurance, 18 March 2026	For small and medium-sized companies, it “can pay for AI-related losses that some General Liability policies exclude”, including bodily injury, property damage and advertising injury. HSB's examples: an AI-controlled HVAC system that creates condensation on the floor, and an AI chat bot's dishwasher installation instructions that cause water damage. Pending regulatory approval at launch, added to the business policies of carriers that partner with HSB.
Coalition	AI affirmative endorsement, 26 March 2024	Added to US surplus and Canada cyber policies; expands a security failure or data breach to include an artificial intelligence security event, and the funds transfer fraud trigger to include deepfake and other AI-transmitted fraudulent instruction.
Relm Insurance	NOVAAI, PONTAAI and RESCAAI, 17 January 2025	NOVAAI covers liability for AI tech products and services, media liability for content produced by AI, AI regulatory liability and deepfake digital crime. PONTAAI is an excess difference in conditions wrap for companies whose current policies may not cover AI-related issues. RESCAAI is for organizations that use third-party AI or embed it in their own products.
Armilla	Armilla Insured and Armilla Guaranteed	Armilla Insured gives “named, affirmative coverage” to AI developers and deployers, backed by Chaucer Group, Axis Capital and Convex. Armilla Guaranteed is an AI performance warranty backed by Chaucer Group, Greenlight Re and Swiss Re. Cover begins with an “AI risk assessment to identify vulnerabilities and compliance gaps”.
Corgi	AI insurance for startups	Addresses “hallucinations, bias claims, and training data disputes”, and says it “can help assemble coverage for AI-related risks in one program”.

Sources: Beazley [14] · HSB [15] · Business Insurance [7] · Reinsurance News [16] · Armilla [17] · Corgi [18]

Market moves

On 19 August 2026 Munich Re Group agreed to acquire At-Bay, a US cyber insurer with gross written premiums of $278 million, at an enterprise value of $575 million, with At-Bay to be overseen by HSB. Munich Re describes a cyber market “rapidly evolving from standalone coverage” towards “integrated, continuously managed risk mitigation platforms.” Closing is expected in the first quarter of 2027.

Source: Munich Re Group [19]

On 5 October 2026 Cowbell launched Risk Advisor, a policyholder-facing AI agent within its OMNI system. For organizations deploying AI, Cowbell says it interprets three AI-specific Cowbell Factors, named AI Exposure, AI Vulnerability and AI Assurance, and prioritizes actions for the policyholder.

Source: Cowbell [20]

Regulator dates for insurers

Date	Regulator	What the source says
4 December 2023	NAIC	Model Bulletin: Use of Artificial Intelligence Systems by Insurers adopted.
11 July 2024	New York DFS	Insurance Circular Letter No. 7, use of AI systems in insurance underwriting.
12 June 2026	Texas Department of Insurance	Bulletin B-0003-26: decisions or actions impacting consumers that are made or supported by AI “must comply with all applicable insurance laws and regulations”. The bulletin also sets out “the type of information and documentation that TDI may request during an investigation or examination.”
27 July 2026	European Union	The AI Omnibus enters into force. Rules for high-risk AI systems in Annex III apply from 2 December 2027; for high-risk AI embedded in products under Annex I, from 2 August 2028.
14 August 2026	Bermuda Monetary Authority	Consultation paper, The Responsible Use of Artificial Intelligence in Financial Services in Bermuda, with a proposed guidance note. Comments to be received by 30 October 2026.
31 August 2026	NAIC	Implementation map: 25 states and the District of Columbia have adopted the Model Bulletin. California, Colorado, New York and Texas are listed under insurance-specific regulation or guidance.
29 September 2026	NAIC Big Data and Artificial Intelligence (H) Working Group	Comment deadline on the AI Risk Evaluation Supplement, version 5.0. Comments received were posted on 2 October 2026. Public meetings to continue discussion are listed for 8 October and 2 November 2026.
2 December 2027	European Union	Annex III high-risk rules apply. Annex III, point 5(c), covers AI systems intended for risk assessment and price setting in relation to natural persons in life and health insurance.

Sources: NAIC [21] · New York Department of Financial Services [22] · Texas Department of Insurance [23] · European Commission [24] · Bermuda Monetary Authority [25] · NAIC [26] · European Commission AI Act Service Desk [27]

Questions the market has not yet answered

The sources above leave these questions open. They are set out as questions, not answers.

Which line responds when an AI agent acts with valid credentials, without a human attacker, and outside what its owner intended?
What evidence will a claim on an AI agent's action require? The AIUC-led report expects AI incidents to need “forensic reconstruction of agentic decision chains”, and claims adjusters who are “technically literate in agentic AI.”
Who authorized the action, and was that authority still in force when the action took effect?
When an insured suspends its own AI system, what will establish when the system went wrong, who ordered the stop, and what ran after it?
How will correlated losses be measured when many insureds depend on the same few foundation model providers?
Will policy language follow claims experience, as the Gallagher survey report expects, or be written before the claims arrive?
With exclusions and affirmative covers side by side, how will an insured know, line by line, whether an agent's loss is covered, excluded or silent?

Sources: Artificial Intelligence Underwriting Company and co-authors [11] · Carrier Management [4]

Sources

Sector research, summarised from published sources for orientation. It is not legal or insurance advice, and it makes no finding on any policy, product or claim. Statements attributed to a company describe that company's own account of its product. To report an error, write to the address in the footer with the page address and the source you believe is authoritative.