Words Speak Louder Than Code: Investigating Cognitive Heuristics in LLM-Based Code Vulnerability Detection

Researchers and practitioners increasingly apply Large Language Models (LLMs) for automated vulnerability detection. Recent work has shown that LLMs are susceptible to the same cognitive heuristics that bias human judgment. Yet, no work has investigated whether these heuristics affect a model's assessment of code vulnerabilities. In this paper, we present the first systematic exploration of cognitive heuristics in LLM-driven code vulnerability detection. We introduce a controlled framework that holds the code fixed and only varies the surrounding context to trigger three cognitive heuristics: the halo effect through author attribution, the framing effect through task objectives and consequences, and the anchoring effect through prior analysis results. Within this framework, we evaluate eight LLMs across three programming languages and perform both quantitative and code-level analyses. Our findings demonstrate that all evaluated models are susceptible to these heuristics. Cross-model average susceptibility is highest for framing at 33.2%, followed by anchoring at 23.5% and halo at 18.4%. Code-level analysis reveals that vulnerabilities that require semantic reasoning for detection are more susceptible to cognitive heuristics than those identifiable through pattern matching. Furthermore, models often change their verdict from safe to vulnerable based on the cognitive condition, without accurately identifying the actual vulnerability. To highlight the practical impact, we demonstrate a proof-of-concept black-box cognitive attack that can suppress up to 97% of previously detected vulnerabilities. These findings indicate that cognitive susceptibility is a consistent and exploitable property of LLM-based vulnerability detection.

RoboJailBench: Benchmarking Adversarial Attacks and Defenses in Embodied Robotic Agents

Recent advances in Vision-Language Models (VLMs) facilitate a new class of embodied AI systems, where these models are integrated into physical platforms, e.g. robots and autonomous vehicles, to interpret visual scenes and execute natural language commands in diverse environments. Previous research has introduced jailbreak attacks and defenses for embodied AI. Their evaluations, however, rely on ad-hoc datasets, limited metrics, and emphasize attack success while neglecting the trade-off between security and the ability to follow benign commands. Existing benchmarks and evaluation frameworks either target traditional chat-based models or focus on non-adversarial safety evaluation for embodied AI; neither captures the adversarial risks, inputs, consequences, and evaluation criteria necessary for jailbreak attacks in embodied AI systems. In this paper, we address this gap with RoboJailBench, which consists of three core components. We establish a security taxonomy derived from ISO standards, regulatory rules, and documented incidents. This effort yields 18 categories of security violation consequences for embodied AI. We introduce an intent contrast dataset pipeline that augments existing datasets with paired adversarial and benign goals to measure both security and utility. Lastly, we provide an evolving repository with standardized metrics and a unified process for assessing and integrating new attacks and defenses. With this benchmark, we construct a new taxonomy-balanced dataset and augment five existing datasets. We integrate four attacks and two defenses to evaluate their performance on leading embodied VLMs. This benchmark provides the first standardized evaluation framework for jailbreak attacks in embodied AI and supports future research. We release our code, datasets, and artifacts, and maintain a leaderboard at https://purseclab.github.io/benchmark-for-robotics-security.

Exploring and Developing a Pre-Model Safeguard with Draft Models

Large Language Model (LLM) alignment remains vulnerable to jailbreak attacks that elicit unsafe responses, motivating pre-model and post-model guards. Pre-model guards audit the safety of prompts before invoking target models. However, relying solely on the prompt often leads to high false-negative rates (i.e., jailbreak attacks go undetected). Post-model guards address this issue by auditing both the user prompt and the target model's response. However, they incur a high computational cost, including increased token usage and processing time, because they operate after target model inference. In this paper, we introduce a safeguard design that leverages the transferability of jailbreak attacks to enforce prompt safety before target model inference. We first conduct a systematic study of jailbreak transferability, particularly from LLMs to small language models (SLMs). Through these experiments, we identify key factors influencing transferability. Building on these insights, we observe that responses from smaller draft models reflect the safety implications of those from large target models; \ie given a jailbreak prompt constructed for an LLM, an SLM is likely to be triggered to generate an unaligned response. Based on this observation, our safeguard design leverages speculative inference with SLMs to generate a set of draft responses. It then feeds the original prompt and these drafts into existing guards to predict their safety. We demonstrate that this design reduces the false-negative rate of pre-model guards and offers a low \Efficiency alternative to post-model guards. \textcolor{red}{\bf Notice: This paper contains examples of harmful language.}

Systematic Discovery of Semantic Attacks in Online Map Construction through Conditional Diffusion

Autonomous vehicles depend on online HD map construction to perceive lane boundaries, dividers, and pedestrian crossings -- safety-critical road elements that directly govern motion planning. While existing pixel perturbation attacks can disrupt the mapping, they can be neutralized by standard adversarial defenses. We present MIRAGE, a framework for systematic discovery of semantic attacks that bypass adversarial defenses and degrade mapping predictions by finding plausible environmental variation (e.g. shadows, wet roads). MIRAGE exploits the latent manifold of real-world data learned by diffusion models, and searches for semantically mutated scenes neighboring the ground truth with the same road topology yet mislead the mapping predictions. We evaluate MIRAGE on nuScenes and demonstrate two attacks: (1) boundary removal, suppressing 57.7% of detections and corrupting 96% of planned trajectories; and (2) boundary injection, the only method that successfully injects fictitious boundaries, while pixel PGD and AdvPatch fail entirely. Both attacks remain potent under various adversarial defenses. We use two independent VLM judges to quantify realism, where MIRAGE passes as realistic 80--84% of the time (vs. 97--99% for clean nuScenes), while AdvPatch only 0--9%. Our findings expose a categorical gap in current adversarial defenses: semantic-level perturbations that manifest as legitimate environmental variation are substantially harder to mitigate than pixel-level perturbations.

Stable GFlowNets with Probabilistic Guarantees

Generative Flow Networks (GFlowNets) learn to sample states proportional to an unnormalized reward. Despite their theoretical promise, practical training is often unstable, exhibiting severe loss spikes and mode collapse. To tackle this, we first assess the sensitivity of GFlowNet objectives, demonstrating that a small Total Variation (TV) distance between the learned and target distributions does not preclude unbounded training loss. Motivated by this mismatch, we establish converse guarantees by deriving loss-to-TV bounds that certify global fidelity from bounded trajectory balance losses. Lastly, we propose Stable GFlowNets, an algorithm that leverages our theoretical results to stabilize training, and empirically demonstrate improved training behavior and superior distributional fidelity.

STARS: Segment-level Token Alignment with Rejection Sampling in Large Language Models

Aligning large language models with human values is crucial for their safe deployment; however, existing methods, such as fine-tuning, are computationally expensive and suboptimal. In contrast, inference-time approaches like Best-of-N sampling require practically infeasible computation to achieve optimal alignment. We propose STARS: Segment-level Token Alignment with Rejection Sampling, a decoding-time algorithm that steers model generation by iteratively sampling, scoring, and rejecting/accepting short, fixed-size token segments. This allows for early correction of the generation path, significantly improving computational efficiency and boosting alignment quality. Across a suite of six LLMs, we show that STARS outperforms Supervised Fine-Tuning (SFT) by up to 14.9 percentage points and Direct Preference Optimization (DPO) by up to 4.3 percentage points on win-rates, while remaining highly competitive with strong Best-of-N baselines. Our work establishes granular, reward-guided sampling as a generalizable, robust, and efficient alternative to traditional fine-tuning and full-sequence ranking methods for aligning LLMs.

Formalizing the Safety, Security, and Functional Properties of Agentic AI Systems

Agentic AI systems, which leverage multiple autonomous agents and Large Language Models (LLMs), are increasingly used to address complex, multi-step tasks. The safety, security, and functionality of these systems are critical, especially in high-stakes applications. However, the current ecosystem of inter-agent communication is fragmented, with protocols such as the Model Context Protocol (MCP) for tool access and the Agent-to-Agent (A2A) protocol for coordination being analyzed in isolation. This fragmentation creates a semantic gap that prevents the rigorous analysis of system properties and introduces risks such as architectural misalignment and exploitable coordination issues. To address these challenges, we introduce a modeling framework for agentic AI systems composed of two foundational models. The first, the host agent model, formalizes the top-level entity that interacts with the user, decomposes tasks, and orchestrates their execution by leveraging external agents and tools. The second, the task lifecycle model, details the states and transitions of individual sub-tasks from creation to completion, providing a fine-grained view of task management and error handling. Together, these models provide a unified semantic framework for reasoning about the behavior of multi-AI agent systems. Grounded in this framework, we define 17 properties for the host agent and 14 for the task lifecycle, categorized into liveness, safety, completeness, and fairness. Expressed in temporal logic, these properties enable formal verification of system behavior, detection of coordination edge cases, and prevention of deadlocks and security vulnerabilities. Through this effort, we introduce the first rigorously grounded, domain-agnostic framework for the systematic analysis, design, and deployment of correct, reliable, and robust agentic AI systems.

A Progressive Transformer for Unifying Binary Code Embedding and Knowledge Transfer

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 structures, 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.

Enhancing LLM-based Autonomous Driving Agents to Mitigate Perception Attacks

There is a growing interest in integrating Large Language Models (LLMs) with autonomous driving (AD) systems. However, AD systems are vulnerable to attacks against their object detection and tracking (ODT) functions. Unfortunately, our evaluation of four recent LLM agents against ODT attacks shows that the attacks are 63.26% successful in causing them to crash or violate traffic rules due to (1) misleading memory modules that provide past experiences for decision making, (2) limitations of prompts in identifying inconsistencies, and (3) reliance on ground truth perception data. In this paper, we introduce Hudson, a driving reasoning agent that extends prior LLM-based driving systems to enable safer decision making during perception attacks while maintaining effectiveness under benign conditions. Hudson achieves this by first instrumenting the AD software to collect real-time perception results and contextual information from the driving scene. This data is then formalized into a domain-specific language (DSL). To guide the LLM in detecting and making safe control decisions during ODT attacks, Hudson translates the DSL into natural language, along with a list of custom attack detection instructions. Following query execution, Hudson analyzes the LLM's control decision to understand its causal reasoning process. We evaluate the effectiveness of Hudson using a proprietary LLM (GPT-4) and two open-source LLMs (Llama and Gemma) in various adversarial driving scenarios. GPT-4, Llama, and Gemma achieve, on average, an attack detection accuracy of 83. 3%, 63. 6%, and 73. 6%. Consequently, they make safe control decisions in 86.4%, 73.9%, and 80% of the attacks. Our results, following the growing interest in integrating LLMs into AD systems, highlight the strengths of LLMs and their potential to detect and mitigate ODT attacks.

Rethinking How to Evaluate Language Model Jailbreak

Large language models (LLMs) have become increasingly integrated with various applications. To ensure that LLMs do not generate unsafe responses, they are aligned with safeguards that specify what content is restricted. However, such alignment can be bypassed to produce prohibited content using a technique commonly referred to as jailbreak. Different systems have been proposed to perform the jailbreak automatically. These systems rely on evaluation methods to determine whether a jailbreak attempt is successful. However, our analysis reveals that current jailbreak evaluation methods have two limitations. (1) Their objectives lack clarity and do not align with the goal of identifying unsafe responses. (2) They oversimplify the jailbreak result as a binary outcome, successful or not. In this paper, we propose three metrics, safeguard violation, informativeness, and relative truthfulness, to evaluate language model jailbreak. Additionally, we demonstrate how these metrics correlate with the goal of different malicious actors. To compute these metrics, we introduce a multifaceted approach that extends the natural language generation evaluation method after preprocessing the response. We evaluate our metrics on a benchmark dataset produced from three malicious intent datasets and three jailbreak systems. The benchmark dataset is labeled by three annotators. We compare our multifaceted approach with three existing jailbreak evaluation methods. Experiments demonstrate that our multifaceted evaluation outperforms existing methods, with F1 scores improving on average by 17% compared to existing baselines. Our findings motivate the need to move away from the binary view of the jailbreak problem and incorporate a more comprehensive evaluation to ensure the safety of the language model.