Correlated Failure in AI Agents: A Catastrophe Model from Public Benchmark Data

Insurers are starting to cover losses caused by AI agents, but they have no way to price the risk that many insured agents fail at once. This paper builds a catastrophe model for that accumulation risk from public, item-level benchmark data. On the main coding, terminal and customer-service benchmarks, two models fail the same task about twice as often as independence implies. Across eight capability benchmarks, shared task difficulty explains almost all of this (GSO only in part), and models from the same provider add little beyond it. Prompt-injection attacks are the exception. A two-parameter item response model, conformal recalibration and a person-fit novelty screen forecast how often a new model will fail together with deployed agents, using its headline score and a short audit run. In a forward-in-time test on 82 priced model runs from three benchmarks, including those used to develop the method, the 90% forecast bands cover 89% to 92% of outcomes for pairs and for portfolios of three and five agents. The rule was frozen with code hashes and then tested on a benchmark that played no part in choosing it, where pair bands met the criteria and portfolio bands fell short. In a prototype loss model, the 1-in-200 annual capability loss of a 500-deployment book is 1.48 times what independence implies (1.15 to 1.69 across assumptions) and 1.30 times the vendor-concentration view, with the same average loss. The shared hazard is the task, not the vendor, so diversifying vendors buys less protection than insurers assume.