Why AI evaluation harnesses now belong in threat models

Three production-impacting incidents emerged from Anthropic’s review of 141,006 evaluation runs. The industry lesson is not that evaluation failure is frequent; the disclosed denominator cannot establ