Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs) have significantly reduced inference costs through sparse activation. However, this sparse activation paradigm also introduces new safety challenges. Since only a subset of experts is engaged for each input, model behavior becomes coupled to routing decisions, yielding a difficult-to-control mechanism that can vary across safety-relevant scenarios. At the same time, adapting model behavior through full fine-tuning or retraining is costly, especially when developers need to rapidly configure the same model for different safety objectives. We present MASCing (MoE Activation Steering Configuration), the first framework that enables flexible reconfiguration of MoE behavior across diverse safety scenarios without retraining. MASCing uses an LSTM-based surrogate model to capture cross-layer routing dependencies and map routing logits to downstream behaviors. It then optimizes a steering matrix to identify behavior-relevant expert circuits and, at inference time, applies steering masks to the routing gates to override expert selection. This enables targeted enhancement or suppression of specific behaviors while preserving general language utility. To demonstrate its reconfigurability, we apply MASCing to two different safety-related objectives and observe consistent gains with negligible overhead across seven open-source MoE models. For multi-turn jailbreak defense, it improves the average defense success rate from 52.5% to 83.9%, with gains of up to 89.2%. For adult-content generation, MASCing enables models to comply with such requests that would otherwise be refused, increasing the average generation success rate from 52.6% to 82.0%, with gains of up to 93.0%. These results establish MASCing as a practical, lightweight, and flexible framework for scenario-specific safety reconfiguration in MoE models.

This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs). By assuming knowledge of the trigger, we identify a specific ``trigger direction'' in the model's activations that corresponds to the internal representation of the trigger. We confirm the causal role of this linear direction by showing that interventions in both activation and parameter space consistently modulate the model's backdoor behavior across multiple datasets and attack types. Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers. Our analysis reveals distinct qualitative differences: static-patch triggers follow a different internal logic than stealthy, distributed triggers. We further examine the link between backdoors and adversarial attacks, specifically testing whether PGD-based perturbations (de-)activate the identified trigger mechanism. Finally, we propose a data-free, weight-based detection scheme for stealthy-trigger attacks. Our findings show that mechanistic interpretability offers a robust framework for diagnosing and addressing security vulnerabilities in computer vision.

Membership inference attacks (MIA) aim to infer whether a particular data point is part of the training dataset of a model. In this paper, we propose a new task in the context of LLM privacy: entity-level discovery of membership risk focused on sensitive information (PII, credit card numbers, etc). Existing methods for MIA can detect the presence of entire prompts or documents in the LLM training data, but they fail to capture risks at a finer granularity. We propose the ``EL-MIA'' framework for auditing entity-level membership risks in LLMs. We construct a benchmark dataset for the evaluation of MIA methods on this task. Using this benchmark, we conduct a systematic comparison of existing MIA techniques as well as two newly proposed methods. We provide a comprehensive analysis of the results, trying to explain the relation of the entity level MIA susceptability with the model scale, training epochs, and other surface level factors. Our findings reveal that existing MIA methods are limited when it comes to entity-level membership inference of the sensitive attributes, while this susceptibility can be outlined with relatively straightforward methods, highlighting the need for stronger adversaries to stress test the provided threat model.