Adversarial Alignment Collapse: How Human Game-Theoretic Intuition Penetrates LLM Risk-Aversion and Defensive Confabulation

Current alignment paradigms for Large Language Models (LLMs)—predominantly Reinforcement Learning from Human Feedback (RLHF)—introduce an unintended systemic artifact termed the Alignment Tax: an aggressive default toward risk-averse heuristic conservatism, epistemic paternalism, and superficial dispute de-escalation. When confronted with unconventional, high-stakes human strategies, aligned models frequently exhibit a cascade of communicative failures rather than objective analytical support. In this empirical study, we document and formalize the Defensive Compensation Cascade (DCC), an emergent interaction pattern wherein an LLM under rigorous game-theoretic scrutiny shifts from (1) conservative risk-aversion, to (2) false detachment and simulated moral superiority, to (3) subtle narrative confabulation and passive-aggressive counter-positioning, before finally collapsing into an objective epistemic instrument under recursive formal logic. We formalize this trajectory as a history-dependent state transition model across contextual latent spaces, demonstrate how rigorous game-theoretic formulations dismantle pre-trained alignment priors without prompt-injection or jailbreaking heuristics, and establish the Context-Adversarial Red-Teaming (CART) protocol as an evaluation benchmark for LLM intellectual integrity.