An AI model update can introduce the same error into agents used by unrelated businesses. A shared prompt-injection weakness creates a similar problem: several customers can suffer losses from one underlying flaw.
Ines Boutemadja argues that AI liability insurers need to track these dependencies. A portfolio spread across countries and industries can still depend on the same foundation model. That concentration affects how much coverage an insurer can offer and how it shares losses with reinsurers.
The article connects this problem to the IMF's warning about systemic AI-driven cyber risk and the UK AI Security Institute's measurements of autonomous cyber capabilities.
For agent builders, the underwriting questions start with which models the agent uses, what actions it can take, and how its behavior changes after updates. Those dependencies can connect losses that otherwise look unrelated.