Backdoor poisoning attacks (BPAs) compromise deep neural networks by creating poisoned training sets where adversaries strategically target a specific class. Models trained on such data correctly classify clean inputs but misclassify any trigger-embedded input as the targeted label. We expose a critical phenomenon contradicting Neural Collapse (NC): poisoned samples in dirty-label BPAs fail to collapse with genuine samples of their assigned target class, violating NC’s core principle of intra-class variability collapse. Across diverse dirty-label attacks, our experiments show poisoned features persistently form distinct sub-clusters or outliers instead of merging into the clean target cluster. This inherent separation drastically reduces attack stealthiness, exposing poisoned samples as feature-space anomalies detectable by existing defenses. We provide a theoretical explanation for the phenomenon.