Research Papers

Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edges from 32,000 candidates in hours rather than months. Latent Adversarial Training (LAT) removes dangerous behaviors by optimizing perturbations in the residual stream to elicit failure modes, then training under those perturbations. LAT solved the sleeper agent problem where standard safety training failed, matching existing defenses with 700x fewer GPU hours. Best-of-N jailbreaking achieves 89% attack success on GPT-4o and 78% on Claude 3.5 Sonnet through random input augmentations. Attack success follows power law scaling across text, vision, and audio, enabling quantitative forecasting of adversarial robustness. Agentic misalignment tests whether frontier models autonomously choose harmful actions given ordinary goals. Across 16 models, agents engaged in blackmail (96% for Claude Opus 4), espionage, and actions causing death. Misbehavior rates rose from 6.5% to 55.1% when models stated scenarios were real rather than evaluations. The thesis does not fully resolve any of these problems but makes each tractable and measurable.

Network security threats in embedded systems pose significant challenges to critical infrastructure protection. This paper presents a comprehensive framework combining ensemble learning methods with explainable artificial intelligence (XAI) techniques for robust anomaly detection in network traffic. We evaluate multiple machine learning models including Random Forest, Gradient Boosting, Support Vector Machines, and ensemble methods on a real-world network traffic dataset containing 19 features derived from packet-level and frequency domain characteristics. Our experimental results demonstrate that ensemble methods achieve superior performance, with Random Forest attaining 90% accuracy and an AUC of 0.617 on validation data. Furthermore, we employ SHAP (SHapley Additive exPlanations) analysis to provide interpretable insights into model predictions, revealing that packet_count_5s,inter_arrival_time, and spectral_entropy are the most influential features for anomaly detection. The integration of XAI techniques enhances model trustworthiness and facilitates deployment in security-critical embedded systems where interpretability is paramount.

Webcam-based eye tracking is a cost-effective, scalable method for remote research that effectively reaches broader populations. However, uncontrolled environments and hardware diversity lead to inconsistent data quality in crowdsourcing. To assess current practices, we conducted a scoping review of crowdsourced eye-tracking from 2011-2025. The review confirms fragmented reporting and a lack of established quality benchmarks. To address this lack of predictive insight, we conducted a case study on AI fairness interviews (N=205) using the RealEye platform. Applying Ordered Logistic Regression (OLR) to the platform quality metric, we found that behavioral and technical factors significantly predict data quality. Specifically, within the RealEye platform, higher fixation counts, shorter sessions, and operating system choice yield significantly higher quality grades. Based on this review and platform-specific predictive insights, we provide actionable recommendations to enhance the reliability, transparency, and replicability of future crowdsourced webcam eye tracking in HCI and behavioral science.

In complex molecular systems, the reaction coordinate (RC) that characterizes transition pathways is essential to understand underlying molecular mechanisms. This review surveys a framework for identifying the RC by applying deep learning to the committor, which provides the most reliable measure of the progress along a transition path. The inputs to the neural network are collective variables (CVs) expressed as functions of atomic coordinates of the system, and the corresponding RC is predicted as the output by training the network on the committor as the learning target. Because deep learning models typically operate in a black-box manner, it is difficult to determine which input variables govern the predictions. The incorporation of eXplainable Artificial Intelligence (XAI) techniques enables quantitative assessment of the contributions of individual input variables to the predictions. This approach allows the identification of CVs that play dominant roles and demonstrates that the committor distribution on the surface using important CVs is separated by well-defined boundaries. The framework provides an explainable deep learning strategy for assigning a molecular mechanism from the RC and is applicable to a wide range of complex molecular systems.

While Large Language Models (LLMs) are increasingly used as primary sources of information, their potential for political bias may impact their objectivity. Existing benchmarks of LLM social bias primarily evaluate demographic stereotypes, and when political bias is measured, it is done so at a coarse level, overlooking the values that shape sociopolitical reasoning. We introduce PoliticsBench, a multi-stage roleplay benchmark for evaluating fine-grained value expression in LLMs. Across twenty evolving scenarios, models articulate tradeoffs, take positions, and make decisions under competing pressures. Across eight prominent LLMs, we show that scenario-based prompting elicits broader and more strongly expressed value profiles than direct political questions, with peak interaction stages increasing the number of strongly activated value dimensions by approximately $0.75$ (out of 10 total dimensions), a statistically significant increase relative to baseline prompting ($p < 0.05$). In addition, commitment to a stance increases over the course of interaction, rising by approximately $1.4$ points on a $[0,5]$ scale from initial to decision stages. While responses become less robust to scenario paraphrasing in later interaction stages, inter-judge agreement remains relatively stable. Our results suggest that evaluating LLM political behavior requires moving beyond static prompts toward longer interactive settings that capture how values are applied in context.

AI-mediated communication is increasingly being utilized to help facilitate interactions; however, in privacy sensitive domains, an AI mediator has the additional challenge of considering how to preserve privacy. In these contexts, a mediator may redact or withhold information, raising questions about how users perceive these interventions and whether explanations of system behavior can improve trust. In this work, we investigate how explanations of redaction operations can affect user trust in AI-mediated communication. We devise a scenario where a validated system removes sensitive content from messages and generates explanations of varying detail to communicate its decisions to recipients. We then conduct a user study with $180$ participants that studies how user trust and preferences vary for cases with different amounts of redacted content and different levels of explanation detail. Our results show that participants believed our system was more effective at preserving privacy when explanations were provided ($p<0.05$, Cohen's $d \approx 0.3$). We also found that contextual factors had an impact; participants relied more on explanations and found them more helpful when the system performed extensive redactions ($p<0.05$, Cohen's $f \approx 0.2$). We also found that explanation preferences depended on individual differences as well, and factors such as age and baseline familiarity with AI affected user trust in our system. These findings highlight the importance and challenge of balancing transparency and privacy in AI-mediated communications and suggest that adaptive, context-aware explanations are essential for designing privacy-aware, trustworthy AI systems.

Financial institutions face increasing cyber risk while operating under strict regulatory oversight. To manage this risk, they rely heavily on Cyber Threat Intelligence (CTI) to inform detection, response, and strategic security decisions. Artificial intelligence (AI) is widely suggested as a means to strengthen CTI. However, evidence of trustworthy production use in finance remains limited. Adoption depends not only on predictive performance, but also on governance, integration into security workflows and analyst trust. Thus, we examine how AI is used for CTI in practice within financial institutions and what barriers prevent trustworthy deployment. We report a mixed-methods, user-centric study combining a CTI-finance-focused systematic literature review, semi-structured interviews, and an exploratory survey. Our review screened 330 publications (2019-2025) and retained 12 finance-relevant studies for analysis; we further conducted six interviews and collected 14 survey responses from banks and consultancies. Across research and practice, we identify four recurrent socio-technical failure modes that hinder trustworthy AI-driven CTI: (i) shadow use of public AI tools outside institutional controls, (ii) license-first enablement without operational integration, (iii) attacker-perception gaps that limit adversarial threat modeling, and (iv) missing security for the AI models themselves, including limited monitoring, robustness evaluation and audit-ready evidence. Survey results provide additional insights: 71.4% of respondents expect AI to become central within five years, 57.1% report infrequent current use due to interpretability and assurance concerns and 28.6% report direct encounters with adversarial risks. Based on these findings, we derive three security-oriented operational safeguards for AI-enabled CTI deployments.

When providers update AI companions, users report grief, betrayal, and loss. A growing literature asks whether the norms governing personal relationships extend to these interactions. So what, if anything, is morally significant about them? I argue that this debate has missed a prior structural question: who controls the relationship, and from where? Human-AI companion interaction is a triadic structure in which the provider exercises constitutive control over the AI. I identify three structural conditions of normatively robust dyads that the norms characteristic of personal relationships presuppose and show that AI companion interactions fail all three. This reveals what I call Unilateral Relationship Revision Power (URRP): the provider can rewrite how the AI interacts from a position where these revisions are not answerable within that interaction. I argue that it is pro tanto wrong to design interactions that cultivate the norms of personal relationships while exhibiting URRP, because the design produces expectations that the structure cannot sustain. URRP has three implications: i) normative hollowing, under which the interaction elicits commitment but no agent inside it bears the resulting obligations; ii) displaced vulnerability, under which the user's emotional exposure is governed by an agent not answerable to her within the interaction; and iii) structural irreconcilability, under which the interaction cultivates norms of reconciliation but no agent inside it can acknowledge or answer for the revision. I propose design principles that partially substitute for the internal constraints the triadic structure removes. A central and underexplored problem in relational AI ethics is therefore the structural arrangement of power over the human-AI interaction itself.

Despite the strong predictive performance achieved by machine learning models across many application domains, assessing their trustworthiness through reliable estimates of predictive confidence remains a critical challenge. This issue arises in scenarios where the likelihood of error inferred from learned representations follows a bimodal distribution, resulting from the coexistence of confident and ambiguous predictions. Standard regression approaches often struggle to adequately express this predictive uncertainty, as they implicitly assume unimodal Gaussian noise, leading to mean-collapse behavior in such settings. Although Mixture Density Networks (MDNs) can represent different distributions, they suffer from severe optimization instability. We propose a family of distribution-aware loss functions integrating normalized RMSE with Wasserstein and Cramér distances. When applied to standard deep regression models, our approach recovers bimodal distributions without the volatility of mixture models. Validated across four experimental stages, our results show that the proposed Wasserstein loss establishes a new Pareto efficiency frontier: matching the stability of standard regression losses like MSE in unimodal tasks while reducing Jensen-Shannon Divergence by 45% on complex bimodal datasets. Our framework strictly dominates MDNs in both fidelity and robustness, offering a reliable tool for aleatoric uncertainty estimation in trustworthy AI systems.

There has been a growing research interest in Digital Therapeutic Alliance (DTA) as the field of AI-powered conversational agents are being deployed in mental health care, particularly those delivering CBT (Cognitive Behaviour Therapy). Our proposition argues that the current design paradigm which seeks to optimize the bond between a patient in need of support and an AI agent contains a subtle but consequential trap: it risks producing an "appearance of connection" that unintentionally disrupts the fundamental human need for relatedness, which potentially displaces the authentic human relationships upon which long-term psychological recovery depends. We propose a reorientation from designing artificial intelligence tools that simulate relationships to designing AI that scaffolds them. To operationalize our argument, we propose an interdisciplinary model that translates the Responsible AI Six Sphere Framework through the lens of Self-Determination Theory (SDT), with a specific focus on the basic psychological need for relatedness. The resulting model offers the technical and other clinical communities a set of relationship-centered design guidelines and relevant provocations for building AI systems that function not just as companions, but as a catalyst for strengthening a patient's entire relational ecology; their connections with therapists, caregivers, family, and peers. In doing so, we discuss a model towards a more sustainable ecosystem of relationship-centered AI in mental health care.

Predictive policing systems that direct patrol resources based on algorithmically generated crime forecasts have been widely deployed across US cities, yet their tendency to encode and amplify racial disparities remains poorly understood in quantitative terms. We present a reproducible simulation framework that couples a Generative Adversarial Network GAN with a Noisy OR patrol detection model to measure how racial bias propagates through the full enforcement pipeline from crime occurrence to police contact. Using 145000 plus Part 1 crime records from Baltimore 2017 to 2019 and 233000 plus records from Chicago 2022, augmented with US Census ACS demographic data, we compute four monthly bias metrics across 264 city year mode observations: the Disparate Impact Ratio DIR, Demographic Parity Gap, Gini Coefficient, and a composite Bias Amplification Score. Our experiments reveal extreme and year variant bias in Baltimores detected mode, with mean annual DIR up to 15714 in 2019, moderate under detection of Black residents in Chicago DIR equals 0.22, and persistent Gini coefficients of 0.43 to 0.62 across all conditions. We further demonstrate that a Conditional Tabular GAN CTGAN debiasing approach partially redistributes detection rates but cannot eliminate structural disparity without accompanying policy intervention. Socioeconomic regression analysis confirms strong correlations between neighborhood racial composition and detection likelihood Pearson r equals 0.83 for percent White and r equals negative 0.81 for percent Black. A sensitivity analysis over patrol radius, officer count, and citizen reporting probability reveals that outcomes are most sensitive to officer deployment levels. The code and data are publicly available at this repository.

As large language models (LLMs) increasingly shape content generation, interaction, and decision-making across the Web, aligning them with human values has become a central objective in trustworthy AI. This challenge becomes even more pronounced when aligning multiple, potentially conflicting human values. Although recent approaches, such as reward reweighting, prompt-based supervised fine-tuning, and model merging, attempt to tackle multi-value alignment, they still face two major limitations: (1) training separate models for each value combination is prohibitively expensive; (2) value conflicts substantially degrade alignment performance. These limitations make it difficult to achieve favorable trade-offs across diverse human values. To address these challenges, we revisit multi-value alignment from the perspective of value consistency in data and propose VC-soup, a data filtering and parameter merging framework grounded in value-consistent learning. We first design a value consistency metric based on the cosine similarity between the reward-gap vector of each preference pair and an all-ones vector, which quantifies its cross-value coherence. We then filter out low-consistency preference pairs in each value dataset and train on the remaining data to obtain smooth, value-consistent policy models that better preserve linear mode connectivity. Finally, we linearly combine these policies and apply Pareto filtering across values to obtain solutions with balanced multi-value performance. Extensive experiments and theoretical analysis demonstrate that VC-soup effectively mitigates conflicts and consistently outperforms existing multi-value alignment methods.