Contrastive steering has been shown as a simple and effective method to adjust the generative behavior of LLMs at inference time. It uses examples of prompt responses with and without a trait to identify a direction in an intermediate activation layer, and then shifts activations in this 1-dimensional subspace. However, despite its growing use in AI safety applications, the robustness of contrastive steering to noisy or adversarial data corruption is poorly understood. We initiate a study of the robustness of this process with respect to corruption of the dataset of examples used to train the steering direction. Our first observation is that contrastive steering is quite robust to a moderate amount of corruption, but unwanted side effects can be clearly and maliciously manifested when a non-trivial fraction of the training data is altered. Second, we analyze the geometry of various types of corruption, and identify some safeguards. Notably, a key step in learning the steering direction involves high-dimensional mean computation, and we show that replacing this step with a recently developed robust mean estimator often mitigates most of the unwanted effects of malicious corruption.

Visual token compression is widely adopted to improve the inference efficiency of Large Vision-Language Models (LVLMs), enabling their deployment in latency-sensitive and resource-constrained scenarios. However, existing work has mainly focused on efficiency and performance, while the security implications of visual token compression remain largely unexplored. In this work, we first reveal that visual token compression substantially degrades the robustness of LVLMs: models that are robust under uncompressed inference become highly vulnerable once compression is enabled. These vulnerabilities are state-specific; failure modes emerge only in the compressed setting and completely disappear when compression is disabled, making them particularly hidden and difficult to diagnose. By analyzing the key stages of the compression process, we identify instability in token importance ranking as the primary cause of this robustness degradation. Small and imperceptible perturbations can significantly alter token rankings, leading the compression mechanism to mistakenly discard task-critical information and ultimately causing model failure. Motivated by this observation, we propose a Compression-Aware Attack to systematically study and exploit this vulnerability. CAA directly targets the token selection mechanism and induces failures exclusively under compressed inference. We further extend this approach to more realistic black-box settings and introduce Transfer CAA, where neither the target model nor the compression configuration is accessible. We further evaluate potential defenses and find that they provide only limited protection. Extensive experiments across models, datasets, and compression methods show that visual token compression significantly undermines robustness, revealing a previously overlooked efficiency-security trade-off.

Behavior Cloning (BC) is a popular framework for training sequential decision policies from expert demonstrations via supervised learning. As these policies are increasingly being deployed in the real world, their robustness and potential vulnerabilities are an important concern. In this work, we perform the first analysis of the efficacy of clean-label backdoor attacks on BC policies. Our backdoor attacks poison a dataset of demonstrations by injecting a visual trigger to create a spurious correlation that can be exploited at test time. We evaluate how policy vulnerability scales with the fraction of poisoned data, the strength of the trigger, and the trigger type. We also introduce a novel entropy-based test-time trigger attack that substantially degrades policy performance by identifying critical states where test-time triggering of the backdoor is expected to be most effective at degrading performance. We empirically demonstrate that BC policies trained on even minimally poisoned datasets exhibit deceptively high, near-baseline task performance despite being highly vulnerable to backdoor trigger attacks during deployment. Our results underscore the urgent need for more research into the robustness of BC policies, particularly as large-scale datasets are increasingly used to train policies for real-world cyber-physical systems. Videos and code are available at https://sites.google.com/view/dataset-poisoning-in-bc.

This paper studies the adversarial robustness of conformal novelty detection. In particular, we focus on two powerful learning-based frameworks that come with finite-sample false discovery rate (FDR) control: one is AdaDetect (by Marandon et al., 2024) that is based on the positive-unlabeled classifier, and the other is a one-class classifier-based approach (by Bates et al., 2023). While they provide rigorous statistical guarantees under benign conditions, their behavior under adversarial perturbations remains underexplored. We first formulate an oracle attack setup, under the AdaDetect formulation, that quantifies the worst-case degradation of FDR, deriving an upper bound that characterizes the statistical cost of attacks. This idealized formulation directly motivates a practical and effective attack scheme that only requires query access to the output labels of both frameworks. Coupling these formulations with two popular and complementary black-box adversarial algorithms, we systematically evaluate the vulnerability of both frameworks on synthetic and real-world datasets. Our results show that adversarial perturbations can significantly increase the FDR while maintaining high detection power, exposing fundamental limitations of current error-controlled novelty detection methods and motivating the development of more robust alternatives.

Assessing the safety of autonomous driving (AD) systems against security threats, particularly backdoor attacks, is a stepping stone for real-world deployment. However, existing works mainly focus on pixel-level triggers that are impractical to deploy in the real world. We address this gap by introducing a novel backdoor attack against the end-to-end AD systems that leverage one or more other vehicles' trajectories as triggers. To generate precise trigger trajectories, we first use temporal logic (TL) specifications to define the behaviors of attacker vehicles. Configurable behavior models are then used to generate these trajectories, which are quantitatively evaluated and iteratively refined based on the TL specifications. We further develop a negative training strategy by incorporating patch trajectories that are similar to triggers but are designated not to activate the backdoor. It enhances the stealthiness of the attack and refines the system's responses to trigger scenarios. Through extensive experiments on 5 offline reinforcement learning (RL) driving agents with 6 trigger patterns and target action combinations, we demonstrate the flexibility and effectiveness of our proposed attack, showing the under-exploration of existing end-to-end AD systems' vulnerabilities to such trajectory-based backdoor attacks.