Mixture-of-experts (MoE) architectures have become the dominant paradigm for scaling open-source large language models, yet their security implications remain insufficiently understood. Recent studies reveal that MoE safety alignment is concentrated in specific ``safety experts'' and ``safety neurons'', but existing methods require modifying model internals, limiting practical applicability. We identify that a small number of critical tokens in harmful prompts disproportionately activate safety experts at specific layers. Motivated by this insight, we propose ClearCut, a prefix-based jailbreak method that exploits this architectural vulnerability through input manipulation alone while remaining mechanistically interpretable. Specifically, ClearCut optimizes a short, semantically natural prefix with a hybrid contrastive-suppression loss that redirects these tokens away from safety experts, bypassing safety alignment without any modification to model weights. Across six mainstream MoE models, ClearCut achieves an average success rate of 64.6\%, outperforming GCG (45.1\%) and AutoDAN (51.2\%), while approaching GateBreaker (61.5\%) which requires model modification. The prefix averages only 5.77 tokens, substantially shorter than existing approaches. By identifying critical tokens and tracking routing changes, ClearCut provides interpretable pathways that reveal how MoE safety mechanisms function. This work exposes a fundamental security risk in MoE architectures, that safety experts can be disproportionately activated and manipulated by a small number of critical tokens. Beyond identifying this vulnerability, we provide ClearCut as a practical tool for security evaluation.
Existing architectural backdoors embed malicious logic directly into model structures to persist after clean training, but they typically rely on handcrafted trigger patterns vulnerable to preprocessing and lack the mechanism to dynamically update attack targets. Addressing these limitations, this paper proposes ArchLock, a dynamic-target architectural backdoor framework designed exclusively for stateful inference deployments that maintain persistent mutable state across requests. ArchLock comprises two core components: a correlation-based statistical trigger detector that utilizes local red-green channel Pearson correlations as the activation signal, thereby reducing dependence on absolute pixel values and exhibiting theoretical invariance to linear transformations; and an Adaptive Confidence Calibration module, which maintains a persistent memory buffer to enable post-deployment target switching through a two-phase protocol. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-100 demonstrate that ArchLock achieves a dynamic target success rate exceeding 90% while preserving clean accuracy. Furthermore, the method exhibits robustness against fine-tuning and pruning and evades three behavioral detectors under the stated stateful deployment assumptions. This work highlights the security risks of mutable architectural components in stateful inference services while explicitly discussing the limitations of the proposed approach under stateless deployment and aggressive preprocessing conditions.