Probabilistic Risk Assessment (PRA) originated in the nuclear safety domain with WASH-1400 (1975), designed to quantify accident risks using empirical failure data. When extended to security threats involving intentional adversaries, it has been argued that traditional PRA assumptions like independent risk inputs, frequentist probability interpretation, and static analysis are fundamentally inappropriate. This paper critically surveys these objections and demonstrates that the identified failures reflect poor analytical practice rather than inherent methodological limitations. We organize adversary modeling uncertainty across four dimensions: adversary objectives, information states, capabilities, and decision-making processes. Using the Ellsberg paradox to distinguish aleatory from epistemic uncertainty, we show how Bayesian methods naturally accommodate the deep uncertainty characteristic of adversarial contexts. We present a unified framework mapping existing security risk methods from Adversarial Risk Analysis (ARA) through game-theoretic equilibrium models, design basis threats, and qualitative scoring systems as successive approximations to a common Bayesian decision-theoretic ideal. Each approximation introduces explicit assumptions that analysts and decision-makers must evaluate against their specific context. This taxonomy enables nuclear security practitioners to make informed methodological choices and supports integrated resource allocation across both safety and security risks.