Search for: All records

Large language models (LLMs) can underpin AI assistants that help users with everyday tasks, such as by making recommendations or performing basic computation. Despite AI assistants’ promise, little is known about the implicit values these assistants display while completing subjective everyday tasks. Humans may consider values like environmentalism, charity, and diversity. To what extent do LLMs exhibit these values in completing everyday tasks? How do they compare with humans? We answer these questions by auditing how six popular LLMs complete 30 everyday tasks, comparing LLMs to each other and to 100 human crowdworkers from the US. We find LLMs often do not align with humans, nor with other LLMs, in the implicit values exhibited. 

Not AvailableDevelopers implement safeguards for large language models (LLMs) to restrict specific responses and prevent the generation of unsafe content. However, a growing number of jailbreak approaches aim to circumvent these safeguards by crafting prompts that elicit prohibited content. The success of these approaches is typically evaluated using techniques such as string matching and natural language understanding, which yield an outcome indicating whether the response violates the safeguard. Our analysis reveals two key limitations in existing jailbreak evaluation methodologies: (1) their evaluation objectives lack clarity and do not directly assess the safety of generated responses, and (2) they conflate different jailbreak response outcomes, oversimplifying the jailbreak result into a binary success/failure outcome. In this paper, we propose a novel set of metrics for evaluating LLM jailbreaks: safeguard violation … 

Cooperative perception (CP) extends detection range and situational awareness in connected and autonomous vehicles by aggregating information from multiple agents. However, attackers can inject fabricated data into shared messages to achieve adversarial attacks. While prior defenses detect object spoofing, object removal attacks remain a serious threat. Nevertheless, prior attacks require unnaturally large perturbations and rely on unrealistic assumptions such as complete knowledge of participant agents, which limits their attack success. In this paper, we present SOMBRA, a stealthy and practical object removal attack exploiting the attentive fusion mechanism in modern CP algorithms. SOMBRA achieves 99% success in both targeted and mass object removal scenarios (a 90%+ improvement over prior art) with less than 1% perturbation strength and no knowledge of benign agents other than the victim. To address the unique vulnerabilities of attentive fusion within CP, we propose LUCIA, a novel trustworthiness-aware attention mechanism that proactively mitigates adversarial features. LUCIA achieves 94.93% success against targeted attacks, reduces mass removal rates by over 90%, restores detection to baseline levels, and lowers defense overhead by 300x compared to prior art. Our contributions set a new state-of-the-art for adversarial attacks and defenses in CP. 

Finding collision-free paths is crucial for autonomous multi-robots (AMRs) to complete assigned missions, ranging from search operations to military tasks. To achieve this, AMRs rely on collaborative collision avoidance algorithms. Unfortunately, the robustness of these algorithms against false data injection attacks (FDIAs) remains unexplored. In this paper, we introduce Raven, a tool to identify effective and stealthy semantic attacks (eg, herding). Effective attacks minimize positional displacement and the number of false data injections by using temporal logic and stochastic optimization techniques. Stealthy attacks remain within sensor noise ranges and maintain spatiotemporal consistency. We evaluate Raven against two state-of-the-art collision avoidance algorithms, ORCA and GLAS. Our results show that a single false data injection impacts multi-robot systems by causing position deviation or even collisions. We evaluate Raven on three testbeds–a numerical simulator, a high-fidelity simulator, and Crazyflie drones. Our results reveal five design flaws in these algorithms and underscore the importance of developing robust defenses against FDIAs. Finally, we propose countermeasures to mitigate the attacks we have uncovered. 

Language model approaches have recently been integrated into binary analysis tasks, such as function similarity detection and function signature recovery. These models typically employ a two-stage training process: pre-training via Masked Language Modeling (MLM) on machine code and fine-tuning for specific tasks. While MLM helps to understand binary code struc- tures, it ignores essential code characteristics, including control and data flow, which negatively affect model generalization. Recent work leverages domain-specific features (e.g., control flow graphs and dynamic execution traces) in transformer-based approaches to improve binary code semantic understanding. However, this approach involves complex feature engineering, a cumbersome and time-consuming process that can introduce predictive uncertainty when dealing with stripped or obfuscated code, leading to a performance drop. In this paper, we introduce PROTST, a novel transformer-based methodology for binary code embedding. PROTST employs a hierarchical training process based on a unique tree-like structure, where knowledge progressively flows from fundamental tasks at the root to more specialized tasks at the leaves. This progressive teacher-student paradigm allows the model to build upon previously learned knowledge, resulting in high-quality embeddings that can be effectively leveraged for diverse downstream binary analysis tasks. The effectiveness of PROTST is evaluated in seven binary analysis tasks, and the results show that PROTST yields an average validation score (F1, MRR, and Recall@1) improvement of 14.8% compared to traditional two-stage training and an average validation score of 10.7% compared to multimodal two-stage frameworks. 

The WebXR API enables immersive AR/VR experiences directly through web browsers on head-mounted displays (HMDs). However, prior research shows that security-sensitive UI properties and the lack of an