Can you insure a software program? Yes, and here is every time it was done, dated

A question asked in public, answered on a page. Under RAND's How Is Artificial Intelligence Insured? (RR-A5130-1, Romanosky and Robinson, 16 September 2026), the lead asked what about insurance for individual agents and their deployment environment, and the author asked back whether a software program can be insured at all, and whether it has ever been done. Can you insure a program?↗, in the Insurance group, is the answer: yes, and the interesting part is what got insured each time. The maker's liability since the 1980s, as technology errors and omissions cover. One named defect in 1997, when Y2K cover was sold with six-figure premiums and then excluded from the standard forms from 1 April 1998. A model's output against a threshold since 2018, Munich Re's first AI policy. The behaviour of a program itself since July 2019, Nexus Mutual's smart-contract cover, with its first payout in February 2020. A warranty on an AI model sold to the buyer in 2024, with the term sheet's exact words. Liability for what the AI did from 2025, and the exclusions filed the same year. And an agent, insured in one deployment, in February 2026. Eight entries, each with the source and the date on it, and a table of what was insured, who held it and what triggered a claim.

The question under the question. The page then answers the lead's own question with the underwriters' words. The AIUC blueprint of July 2026, whose listed authors include Sasha Romanosky, says a certified agent executing trades in a brokerage and a certified agent summarising documents for a law firm may share a model and guardrails while their exposure diverges enormously, and lists what an underwriter must ask: does it recommend or execute, what are its access-control configurations, what human oversight and handoff, what least-privilege scoping of the tools and data it reaches. A table maps each question to the part of an Agent Behaviour Policy that answers it. The strategy section says where we come in: not a carrier, not a broker; the three rungs; the bottom rung on sale now, as sixteen templates and the prompt that lets any agent map its own grant.

What the page does not do. It names carriers and products with what their own releases say and when, and nothing about how good they are. Two pages refused a direct read, RAND's own and the Justia page for Winter, and the page says which claims rest on what was read instead. The carrier and limit behind the first insured agent deployment are not public and are not claimed. Dev moved to v1.26.0 while this was being written, so the branch merged it and regenerated; the new page sits in the five-entry menu as built there.