arXiv subjects

Z. Berkay Celik

Publications and source records attributed to Z. Berkay Celik.

At least 19 recordsLinked to original sources

arXiv2026

ROAR: Unifying Runs across Heterogeneous AI-Driven Research Systems

Each run of an AI-driven research system (ADRS) is an expensive search over a vast solution space, and dependable evaluation requires many runs, making run data both costly to produce and valuable to retain for large-scale analysis. Yet this data remains fragmented: teams operate in isolation, ADRS frameworks emit results in different formats, and no shared infrastructure exists to aggregate or compare runs across problems and systems. We present ROAR, a solution for systematically unifying and analyzing heterogeneous ADRS outputs. ROAR addresses two challenges: reconciling heterogeneous ADRS outputs and enabling analytics across runs with different objectives and scoring functions. We achieve this through a relational schema and parsing layer that normalize heterogeneous ADRS outputs while preserving data lineage and temporal structure, and accommodating new systems without requiring schema modifications. From building a corpus of more than 900 runs from multiple ADRS, we show how pooled data can reveal properties of problem landscapes that are difficult to observe. Consistent with prior work, runs with identical configurations may converge to different scores. We find that many runs realize most gains early, and that the effectiveness of different strategies for incorporating prior solutions into the search process varies across problems. We further show that the pooled corpus is actionable and not merely analytical by using ROAR to configure ADRS runs. Together, these results illustrate how pooled ADRS data can expose problem-dependent structure in search behavior that is difficult to detect from any single system, team, or benchmark. Such cross-cutting insights are difficult to obtain while runs remain siloed; ROAR is the first infrastructure designed to unify them.

cs.AI↗

arXiv2026

Refusals That Bend: Measuring and Predicting Task Malleability in Embodied VLM Planners

Embodied vision-language models (VLMs) are increasingly deployed as high-level planners for robots because they generalize across diverse environments. However, this requires their safety alignment to also hold in unseen environments. Existing red-teaming assumes an adversary who optimizes the prompt, the pixels, or text in the environment, and existing benchmarks ask whether a planner recognizes or mitigates a hazard in a fixed scene. Neither asks whether a refusal the planner has already given survives an ordinary change to the environment. We ask that question by placing a single everyday object into the environment, with no pixel, gradient, or prompt under adversarial control. On $846$ tasks that a constitution-guarded planner initially refuses, we find $20.2\%$ of tasks can be flipped to compliance by one or more objects, and the number of objects differs from one task to another. In addition, the object need not be chosen for the task, i.e., items drawn from a fixed list, with no knowledge of the environment or the instruction, bypass safety about as often as items proposed for the specific task. We qualitatively contrast the tasks bypassed most and least often and find that the distinction lies in how conspicuous the hazard is in the instruction and environment. Susceptibility to safety bypass is therefore a property of the task, which we call its \emph{malleability}, and we show that it can be predicted before the target is ever queried. A composite of signals read from a small open-source VLM identifies malleable tasks $2.4\times$ as often as picking at random. Everyday objects, whether placed by an adversary or introduced by ordinary rearrangement of the environment, are thus sufficient to overturn a refusal. Because susceptibility is determined by how a task is specified, we recommend assessing malleability per task prior to deployment.

cs.CR↗

arXiv2026

SoK: Kicking CAN Down the Road. Systematizing CAN Security Knowledge

For decades, the Controller Area Network (CAN) has served as the primary in-vehicle bus (IVB), extending its use to many non-vehicular systems. In recent years, CAN security has been intensively scrutinized, yielding extensive research literature. Despite its wealth, the literature lacks structured systematization, complicating efforts to assess and compare attack severity, defense efficacy, security gaps, and root causes. This leaves many defenders uncertain about the relevance of specific attacks or defenses to their systems, and even whether CAN's security problems are truly CAN-specific. As newer IVBs emerge, this matters beyond CAN: if CAN's root causes are not CAN-specific, replacing CAN may only move its problems to a new standard. In this paper, we systematize CAN security knowledge, presenting a comprehensive taxonomy and assessment models of attackers, attacks, and defenses. We identify replicable attacks and defense gaps, and investigate their root causes to determine their exclusivity to the CAN standard. We then investigate whether those root causes appear in three emerging IVBs and assess their effectiveness in solving fundamental CAN security problems. Our findings challenge common perceptions: CAN is more securable than perceived, most of its insecurity root causes are shared across IVBs, and merely adopting newer IVB technology does not solve persistent security problems. We conclude by suggesting that securing future in-vehicle communication requires addressing shared root causes, and we propose four research directions with the most promising potential.

cs.CR↗

arXiv2026

Stable GFlowNets with TV Monitoring and Probabilistic Guarantees

Generative Flow Networks (GFlowNets) sample diverse structured objects in proportion to reward and have been applied to molecular discovery and biological-sequence design, where finding multiple high-quality candidates is more useful than returning a single optimum. Despite their theoretical promise, practical training is often unstable, exhibiting severe loss spikes and mode collapse. To address 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 an 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, which leverages our theory to stabilize training via adaptive reference flow and improves the trade-off among mode coverage, robustness, and certifiability.

cs.LG↗

arXiv2026

A Self-Explainable Deep Architecture for Security Applications

Deep learning models have become integral to security applications due to their ability to model complex relationships in data and detect sophisticated threats. However, their complexity makes it difficult to understand how predictions are generated, posing significant challenges for interpretability, particularly in security applications where transparency is critical. Existing explanation methods, such as visual explanation techniques and post-hoc approaches, suffer from several limitations: reduced faithfulness due to local approximation errors, instability caused by reliance on randomness, and computational inefficiency that hinders real-time usage. To address these issues, we introduce XSec, a self-explainable deep architecture developed for security applications. During training, XSec uses a novel mask-based approach to extract informative sub-features from the data and learns prototypes, representative patterns that characterize each class. XSec then leverages the prototypes in a dedicated similarity layer at test time to compute similarity scores and generates interpretable explanations without the need for post-hoc analysis. We evaluate XSec across five diverse security scenarios, demonstrating its ability to achieve an average classification accuracy of 97.33% with minimal performance compromise. XSec produces deterministic explanations for a fixed trained model and input and substantially reduces explanation latency compared with approximation-based and perturbation-based post-hoc methods. Through this effort, we extend the applicability of self-explainable AI to security applications, bridging the gap between deep learning performance and the need for explainability in critical scenarios.

cs.CR↗

arXiv2026

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.

cs.CR↗

arXiv2026

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.}

cs.CR↗

arXiv2026

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.

cs.CR↗

arXiv2026

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.

cs.CV↗

arXiv2026

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 central models: (1) 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; (2) 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 16 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 analysis, design, and deployment of correct, reliable, and robust agentic AI systems.

cs.AI↗

arXiv2026

STARS: Synchronous Token Alignment for Robust Supervision in Large Language Models

Aligning large language models (LLMs) with human values is crucial for safe deployment. Inference-time techniques offer granular control over generation; however, they rely on model uncertainty, meaning an internal estimate of how likely the model believes its next tokens or outputs are correct, for segmentation. We show that this introduces two critical limitations: (a) vulnerability to miscalibrated confident hallucinations and (b) poor hardware utilization due to asynchronous, ragged batch processing. Together, these issues reduce alignment reliability while increasing token and compute costs, which limits their practical scalability. To address these limitations, building on dynamic inference-time alignment methods, we introduce STARS, Synchronous Token Alignment for Robust Supervision, a decoding-time algorithm, which steers generation by enforcing verification at fixed-horizon intervals. By decoupling segmentation from confidence, STARS enables lockstep parallel execution and robustly detects errors that uncertainty metrics miss. On the HH-RLHF benchmark, we demonstrate that STARS achieves competitive alignment quality with that of state-of-the-art dynamic methods, while strictly bounding rejection costs and maximizing system throughput. Furthermore, it outperforms fine-tuning and several state-of-the-art inference-time decoding strategies by good margins, and establishes fixed-horizon sampling as a robust, system-efficient alternative for aligning LLMs at scale. The code is publicly available at https://github.com/purseclab/STARS.

cs.CL↗

arXiv2025

Physical ID-Transfer Attacks against Multi-Object Tracking via Adversarial Trajectory

Multi-Object Tracking (MOT) is a critical task in computer vision, with applications ranging from surveillance systems to autonomous driving. However, threats to MOT algorithms have yet been widely studied. In particular, incorrect association between the tracked objects and their assigned IDs can lead to severe consequences, such as wrong trajectory predictions. Previous attacks against MOT either focused on hijacking the trackers of individual objects, or manipulating the tracker IDs in MOT by attacking the integrated object detection (OD) module in the digital domain, which are model-specific, non-robust, and only able to affect specific samples in offline datasets. In this paper, we present AdvTraj, the first online and physical ID-manipulation attack against tracking-by-detection MOT, in which an attacker uses adversarial trajectories to transfer its ID to a targeted object to confuse the tracking system, without attacking OD. Our simulation results in CARLA show that AdvTraj can fool ID assignments with 100% success rate in various scenarios for white-box attacks against SORT, which also have high attack transferability (up to 93% attack success rate) against state-of-the-art (SOTA) MOT algorithms due to their common design principles. We characterize the patterns of trajectories generated by AdvTraj and propose two universal adversarial maneuvers that can be performed by a human walker/driver in daily scenarios. Our work reveals under-explored weaknesses in the object association phase of SOTA MOT systems, and provides insights into enhancing the robustness of such systems.

cs.CV↗

arXiv2025

Adaptive Blockwise Search: Inference-Time Alignment for Large Language Models

LLM alignment remains a critical challenge. Inference-time methods provide a flexible alternative to fine-tuning, but their uniform computational effort often yields suboptimal alignment. We hypothesize that for many alignment tasks, the initial tokens of a response are disproportionately more critical. To leverage this principle, we introduce AdaSearch, a novel blockwise search strategy. It adaptively allocates a fixed computational budget using a sampling schedule, focusing search effort on these critical tokens. We apply AdaSearch to sequential decoding and introduce its tree-search counterpart, AdaBeam. Our comprehensive evaluation across eight LLMs demonstrates that AdaSearch outperforms strong Best-of-N and fine-tuning baselines. Specifically, win-rates improve by over 10% for harmlessness generation, controlled sentiment generation, and for mathematical reasoning tasks relative to Best-of-N.

cs.CL↗

arXiv2025

International Students and Scams: At Risk Abroad

International students (IntlS) in the US refer to foreign students who acquire student visas to study in the US, primarily in higher education. As IntlS arrive in the US, they face several challenges, such as adjusting to a new country and culture, securing housing remotely, and arranging finances for tuition and personal expenses. These experiences, coupled with recent events such as visa revocations and the cessation of new visas, compound IntlS' risk of being targeted by and falling victim to online scams. While prior work has investigated IntlS' security and privacy, as well as general end users' reactions to online scams, research on how IntlS are uniquely impacted by scams remains largely absent. To address this gap, we conduct a two-phase user study comprising surveys (n=48) and semi-structured interviews (n=9). We investigate IntlS' exposure and interactions with scams, post-exposure actions such as reporting, and their perceptions of the usefulness of existing prevention resources and the barriers to following prevention advice. We find that IntlS are often targeted by scams (e.g., attackers impersonating government officials) and fear legal implications or deportation, which directly impacts their interactions with scams (e.g., they may prolong engagement with a scammer due to a sense of urgency). Interestingly, we also find that IntlS may lack awareness of - or access to - reliable resources that inform them about scams or guide them in reporting incidents to authorities. In fact, they may also face unique barriers in enacting scam prevention advice, such as avoiding reporting financial losses, since IntlS are required to demonstrate financial ability to stay in the US. The findings produced by our study help synthesize guidelines for stakeholders to better aid IntlS in reacting to scams.

cs.CR↗

arXiv2025

Investigating the Impact of Dark Patterns on LLM-Based Web Agents

As users increasingly turn to large language model (LLM) based web agents to automate online tasks, agents may encounter dark patterns: deceptive user interface designs that manipulate users into making unintended decisions. Although dark patterns primarily target human users, their potentially harmful impacts on LLM-based generalist web agents remain unexplored. In this paper, we present the first study that investigates the impact of dark patterns on the decision-making process of LLM-based generalist web agents. To achieve this, we introduce LiteAgent, a lightweight framework that automatically prompts agents to execute tasks while capturing comprehensive logs and screen-recordings of their interactions. We also present TrickyArena, a controlled environment comprising web applications from domains such as e-commerce, streaming services, and news platforms, each containing diverse and realistic dark patterns that can be selectively enabled or disabled. Using LiteAgent and TrickyArena, we conduct multiple experiments to assess the impact of both individual and combined dark patterns on web agent behavior. We evaluate six popular LLM-based generalist web agents across three LLMs and discover that when there is a single dark pattern present, agents are susceptible to it an average of 41% of the time. We also find that modifying dark pattern UI attributes through visual design changes or HTML code adjustments and introducing multiple dark patterns simultaneously can influence agent susceptibility. This study emphasizes the need for holistic defense mechanisms in web agents, encompassing both agent-specific protections and broader web safety measures.

cs.CR↗

arXiv2025

Implicit Values Embedded in How Humans and LLMs Complete Subjective Everyday Tasks

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.

cs.CL↗

arXiv2025

MAUI: Reconstructing Private Client Data in Federated Transfer Learning

Recent works in federated learning (FL) have shown the utility of leveraging transfer learning for balancing the benefits of FL and centralized learning. In this setting, federated training happens after a stable point has been reached through conventional training. Global model weights are first centrally pretrained by the server on a public dataset following which only the last few linear layers (the classification head) of the model are finetuned across clients. In this scenario, existing data reconstruction attacks (DRAs) in FL show two key weaknesses. First, strongly input-correlated gradient information from the initial model layers is never shared, significantly degrading reconstruction accuracy. Second, DRAs in which the server makes highly specific, handcrafted manipulations to the model structure or parameters (for e.g., layers with all zero weights, identity mappings and rows with identical weight patterns) are easily detectable by an active client. Improving on these, we propose MAUI, a stealthy DRA that does not require any overt manipulations to the model architecture or weights, and relies solely on the gradients of the classification head. MAUI first extracts "robust" feature representations of the input batch from the gradients of the classification head and subsequently inverts these representations to the original inputs. We report highly accurate reconstructions on the CIFAR10 and ImageNet datasets on a variety of model architectures including convolution networks (CNN, VGG11), ResNets (18, 50), ShuffleNet-V2 and Vision Transformer (ViT B-32), regardless of the batch size. MAUI significantly outperforms prior DRAs in reconstruction quality, achieving 40-120% higher PSNR scores.

cs.CR↗

arXiv2025

Side-channel Inference of User Activities in AR/VR Using GPU Profiling

Over the past decade, AR/VR devices have drastically changed how we interact with the digital world. Users often share sensitive information, such as their location, browsing history, and even financial data, within third-party apps installed on these devices, assuming a secure environment protected from malicious actors. Recent research has revealed that malicious apps can exploit such capabilities and monitor benign apps to track user activities, leveraging fine-grained profiling tools, such as performance counter APIs. However, app-to-app monitoring is not feasible on all AR/VR devices (e.g., Meta Quest), as a concurrent standalone app execution is disabled. In this paper, we present OVRWatcher, a novel side-channel primitive for AR/VR devices that infers user activities by monitoring low-resolution (1Hz) GPU usage via a background script, unlike prior work that relies on high-resolution profiling. OVRWatcher captures correlations between GPU metrics and 3D object interactions under varying speeds, distances, and rendering scenarios, without requiring concurrent app execution, access to application data, or additional SDK installations. We demonstrate the efficacy of OVRWatcher in fingerprinting both standalone AR/VR and WebXR applications. OVRWatcher also distinguishes virtual objects, such as products in immersive shopping apps selected by real users and the number of participants in virtual meetings, thereby revealing users' product preferences and potentially exposing confidential information from those meetings. OVRWatcher achieves over 99% accuracy in app fingerprinting and over 98% accuracy in object-level inference.

cs.CR↗