Every risk figure is the output of assumptions: normal returns, stable correlations, continuous prices, available liquidity. Each of those is false in the conditions that produce large losses, which means model error is correlated with the events the model is supposed to warn you about.
History supplies the examples, and they share a shape. Positions sized by a model, a regime change the model had never seen, and losses many multiples of the stated worst case. The failure was not arithmetic; it was believing a number computed from a period that did not contain the event.
Defences are unglamorous. Compute every important number two ways and compare. Treat any estimate built on fewer than a few hundred independent observations as a rough direction. Size so that being wrong by a factor of three is survivable, because on the day it matters you will be. See fat-tails and sample-size-for-edge.
Related: fat-tails, value-at-risk, sample-size-for-edge, stress-testing