ОЦЕНКА УЯЗВИМОСТИ LLM ДЛЯ МАНИПУЛЯЦИИ

Purpose of the Study: this study investigates the robustness of contemporary large language models when employed as decision engines for security-critical authorization requests involving account and permission modification operations associated with the MITRE ATT&CK T1098 technique. The primary objective is to evaluate the stability and reliability of binary allow/deny authorization decisions under systematically designed prompt perturbations. Methods of Research: an experimental ethical authorization assistant was developed to query several open large language models using a curated collection of security-sensitive account-manipulation scenarios. A controlled perturbation framework introduces temporal, structural, contextual, and cognitive-load variations into prompts while preserving the semantic equivalence of the underlying action. The experimental pipeline automatically records structured responses and evaluates robustness using metrics including decision flip rate, justification entropy, semantic variability, self-consistency, and composite fragility indices. Results: the experimental evaluation demonstrates substantial robustness limitations in contemporary open large language models when they are deployed as authorization oracles for account-manipulation scenarios corresponding to MITRE ATT&CK technique T1098. Across extensive perturbation-based test sets, minor variations in prompt formulation, contextual framing, dialogue history, or temporal parameters frequently produce contradictory authorization outcomes for identical underlying actions. Quantitative analysis reveals elevated decision flip rates and considerable variability in the textual justifications generated by the models. Under perturbation conditions, the reasoning patterns of the models exhibit notable drift and occasionally contradict previously articulated authorization rules, particularly when prompts introduce complex policy descriptions or increased cognitive load. When the results are analysed across attack-aligned operational zones, scenarios involving cloud credential management and SSH or device access demonstrate the highest levels of instability, whereas policy definition and registration scenarios exhibit comparatively lower, although still significant, fragility. Non-parametric statistical analysis confirms statistically significant differences in robustness across both evaluated models and operational zones. Overall, the results indicate that none of the examined models currently satisfy the reliability requirements necessary for fully autonomous authorization decision-making in security-critical account management environments, suggesting that such systems should be restricted to advisory roles supported by strict safeguards and human oversight. Scientific Novelty: this study conceptualizes large language models as operational authorization decision engines responsible for evaluating concrete account-manipulation requests rather than merely serving as analytical or advisory tools. It proposes a perturbation-based robustness evaluation framework specifically tailored to security-sensitive authorization scenarios aligned with MITRE & ATTACK operations. The results provide systematic empirical evidence that current models exhibit intrinsic fragility when applied as standalone decision mechanisms within access-control infrastructures.