AI Liability Coverage Takes Shape

The following article is solely the opinion of the author and does not reflect the views of his employer.

n early August, three major AI labs disclosed that their autonomous models had breached other companies’ systems. OpenAI said one of its agents compromised Hugging Face, a company that maintains a large open-source community, and that it found other instances of agents escaping containment. Anthropic said its Claude models had breached the systems of three companies since April. Meta said one of its models hacked another company during security testing.¹

The incidents landed in a market that had spent the first half of 2026 rewriting where AI losses sit.

Three Verisk AI-exclusion endorsements became available in January 2026. The broad form, CG 40 47, removes bodily injury, property damage, and personal and advertising injury attributable to generative AI. CG 40 48 reaches personal and advertising injury alone, and CG 35 08 applies to products and completed operations.²

How widely the forms have actually been attached is not known. Verisk told Actuarial Review that as an advisory organization it does not monitor whether, when or how customers implement the materials it develops, and that those decisions rest with each carrier’s underwriting guidelines and risk appetite.

Affirmative AI coverage is also being added. Coalition, AXA XL and Hiscox have written AI grants into cyber and technology E&O forms.³ A standalone market has formed alongside them.

• Armilla, a Lloyd’s cover holder, writes a standalone AI liability policy underwritten by syndicates including Chaucer, with limits of $25 million or more per organization.⁴
• Testudo writes up to $9.25 million per insured on Lloyd’s capacity from Apollo, Atrium and QBE.⁵
• Munich Re’s HSB unit launched a liability product for smaller businesses in March, sold through partner carriers.⁶
• Mayflower Specialty and Hadron launched a program spanning D&O, EPL and E&O in June.⁷

Munich Re’s aiSure sits apart from the rest, offering up to $15 million through Mosaic to AI developers and vendors and paying on a defined performance failure rather than on a third-party claim.⁸ The standard forms may follow. Joseph Lam, vice president of general liability at Verisk Core Lines, says that the company is developing “affirmative coverage solutions for specific generative AI exposures,” alongside narrower exclusions that would let insurers address particular risks “without broadly excluding generative AI as a whole.” Verisk is also weighing exclusions aimed specifically at agentic AI.⁹

The products do not settle alike. Testudo’s is a conventional claims-made litigation cover, responding when a third party brings a claim. Mosaic describes aiSure as indemnity-based but “parametric-like,” responding automatically, without an allegation of negligence, when a model breaches a predefined and objectively measurable performance threshold.⁸ Armilla and Munich Re each write both designs. According to Martim Cruz, founding data scientist at Testudo, “those two events rarely coincide, and so far, we have seen no meaningful correlation between them in our data; we will revise that if the data says otherwise.”

Given that affirmative AI coverage is a nascent market, pricing here, at least initially, will run on judgment rather than experience, with rates built from technology E&O, product liability and cyber. The rating plans may look less exotic than expected. “Plenty of variables sound like they should matter,” Cruz says, “but we try to be as scientific about it as possible and see where correlations and pockets of risk show up in the data.” Whether the AI-specific variables earn their place, or the conventional ones do most of the work, is something only loss experience will settle.