From biotechnology to cyber-risks, most extreme technological risks cannot be reliably estimated from historical statistics. Therefore, engineers resort to predictive methods, such as fault/event trees in the framework of probabilistic safety assessment (PSA), which consists in developing models to identify triggering events, potential accident scenarios, and estimate their severity and frequency. However, even the best safety analysis struggles to account for evolving risks resulting from inter-connected networks and cascade effects. Taking nuclear risks as an example, the predicted plant-specific distribution of losses is found to be significantly underestimated when compared with available empirical records. Using a novel database of 99 events with losses larger than $50'000 constructed by Sovacool, we document a robust power law distribution with tail exponent mu \approx 0.7. A simple cascade model suggests that the classification of the different possible safety regimes is intrinsically unstable in the presence of cascades. Additional continuous development and validation, making the best use of the experienced realized incidents, near misses and accidents, is urgently needed to address the existing known limitations of PSA when aiming at the estimation of total risks.