As large language models (LLMs) are increasingly deployed for complex reasoning tasks, Long Chain-of-Thought (Long-CoT) prompting has emerged as a key paradigm for structured inference. Despite early-stage safeguards enabled by alignment techniques such as RLHF, we identify a previously underexplored vulnerability: reasoning trajectories in Long-CoT models can drift from aligned paths, resulting in content that violates safety constraints. We term this phenomenon Path Drift. Through empirical analysis, we uncover three behavioral triggers of Path Drift: (1) first-person commitments that induce goal-driven reasoning that delays refusal signals; (2) ethical evaporation, where surface-level disclaimers bypass alignment checkpoints; (3) condition chain escalation, where layered cues progressively steer models toward unsafe completions. Building on these insights, we introduce a three-stage Path Drift Induction Framework comprising cognitive load amplification, self-role priming, and condition chain hijacking. Each stage independently reduces refusal rates, while their combination further compounds the effect. To mitigate these risks, we propose a path-level defense strategy incorporating role attribution correction and metacognitive reflection (reflective safety cues). Our findings highlight the need for trajectory-level alignment oversight in long-form reasoning beyond token-level alignment.

UNLABELLED: SORTALLER is an online allergen classifier based on allergen family featured peptide (AFFP) dataset and normalized BLAST E-values, which establish the featured vectors for support vector machine (SVM). AFFPs are allergen-specific peptides panned from irredundant allergens and harbor perfect information with noise fragments eliminated because of their similarity to non-allergens. SORTALLER performed significantly better than other existing software and reached a perfect balance with high specificity (98.4%) and sensitivity (98.6%) for discriminating allergenic proteins from several independent datasets of protein sequences of diverse sources, also highlighting with the Matthews correlation coefficient (MCC) as high as 0.970, fast running speed and rapidly predicting a batch of amino acid sequences with a single click. AVAILABILITY AND IMPLEMENTATION: http://sortaller.gzhmc.edu.cn/.