Research Papers

Evaluating Large Language Models in a Complex Hidden Role Game

Niklas Bauer

Quantifying the deceptive potential of Large Language Models (LLMs) is critical for AI safety, yet difficult to achieve in uncontrolled environments. This work investigates the reasoning, persuasion, and deceptive capabilities of LLMs within the social deduction game Secret Hitler. I introduce an open-source framework and novel metrics to measure performance: Role Identification Accuracy, Deception Retention Rate, and Game State Impact Rate. By benchmarking models against rule-based algorithms and human games, I identify a gap between conversational ability and strategic depth. The study also analyzes the impact of reasoning-enhancement techniques on win rates and strategic reasoning. Neither Chain-of-Thought prompting nor internal memory bring improvements in performance, with up to 23.2% worse win rates for fascist roles. While rule-based agents align with expert human voting decisions 86.7% of the time, models like Llama 3.1 70B achieve only a 59.7% accuracy. Models playing as Fascists consistently yield negative impact scores and fail to sustain deception, resulting in roughly 40% shorter games compared to humans. These findings suggest that current architectures remain ineffective at complex, multi-turn manipulation. As capabilities advance, detecting when models begin to master these deceptive behaviors is crucial. The developed framework serves as a reproducible testbed for future alignment research.

This study examines the perception of legal professionals on the governance of AI in developing countries, using Nigeria as a case study. The study focused on ethical risks, regulatory gaps, and institutional readiness. The study adopted a qualitative case study design. Data were collected through 27 semi-structured interviews with legal practitioners in Nigeria. A focus group discussion was also held with seven additional legal practitioners across sectors such as finance, insurance, and corporate law. Thematic analysis was employed to identify key patterns in participant responses. Findings showed that there were concerns about data privacy risks and the lack of enforceable legal frameworks. Participants expressed limited confidence in institutional capacity and emphasized the need for locally adapted governance models rather than direct adoption of foreign frameworks. While some expressed optimism about AI's potential, this was conditional on the presence of strong legal oversight and public accountability. The study contributes to the growing discourse on AI governance in developing countries by focusing on the perspectives of legal professionals. It highlights the importance of regulatory approaches that are context-specific, inclusive, and capable of bridging the gap between global ethical principles and local realities. These insights offer practical guidance for policymakers, regulators, and scholars working to shape responsible AI governance in similar environments.

The rapid advancement of Large Language Models (LLMs) in biological research has significantly lowered the barrier to accessing complex bioinformatics knowledge, ex perimental design strategies, and analytical workflows. While these capabilities accelerate innovation, they also introduce serious dual-use risks, as Bio-LLMs can be exploited to generate harmful biological insights under the guise of legitimate research queries. Existing safeguards, such as static prompt filtering and policy-based restrictions, are insufficient when LLMs are embedded within dynamic biological workflows and application-layer systems. In this paper, we present BioShield, a context-aware application-level firewall designed to secure Bio LLMs against dual-use attacks. At the core of BioShield is a domain-specific prompt scanner that performs contextual risk analysis of incoming queries. The scanner leverages a harmful scoring mechanism tailored to biological dual-use threat cat egories to identify prompts that attempt to conceal malicious intent within seemingly benign research requests. Queries ex ceeding a predefined risk threshold are blocked before reaching the model, effectively preventing unsafe knowledge generation at the source. In addition to pre-generation protection, BioShield deploys a post-generation output verification module that inspects model responses for actionable or weaponizable biological content. If an unsafe response is detected, the system triggers controlled regeneration under strengthened safety constraints. By combining contextual prompt scanning with response-level validation, BioShield provides a layered defense framework specifically designed for bio-domain LLM deployments. Our framework advances cyberbiosecurity by formalizing dual-use threat detection in Bio-LLMs and proposing a structured mitigation strategy for secure, responsible AI driven biological research.

As generative AI technologies rapidly advance, AI agents are gaining the ability not only to collect data and perform tasks but also to respond to environments and evolve over time. This shift opens new possibilities for reimagining digital legacy - raising critical questions about how we remember, commemorate, and interact with the traces of the deceased. The forms of these AI agents are particularly important, as they act as vessels for digital legacies - much like urns for ashes. We will ask: What kinds of devices or representations would we want to store our digital selves or legacies in? How do we envision future generations interacting with these forms? The question is not only about the function of these agents or the object's role as a storage vessel but also the meaning it carries, the memories it preserves, and its connection to the extended notion of our "Generative Horcrux." This three-hour in-person workshop invites design practitioners and researchers from diverse backgrounds to explore the emerging landscape of generative AI agent-based digital legacy. This workshop uses fiction and hands on prototyping to explore how AI agents might reconfigure memory, identity, and posthumous presence in future sociotechnical worlds. We anticipate that this session will foster interdisciplinary dialogue and contribute conceptually and methodologically to HCI, design research, and AI ethics.

Generative Artificial Intelligence (GAI) offers new opportunities for reconstructing these unrecorded memory scenes, yet existing web-based tools undermine users' sense of agency through disengaging and unpredictable interactions. In this work, we advance three design arguments about how slow, tangible interaction can reshape human-AI relationships by making temporality, embodied agency, and generative processes experientially legible. We instantiate these arguments by presenting Memory Printer, a tangible design that combines silk-screen printing metaphors with text-to-image generation. The design features layered reconstruction that decomposes image generation into incremental steps, a physical wooden scraper enabling embodied control over image revelation, and built-in printing that produces tangible photos. We examine these arguments through a comparative study with 24 participants, exploring how participants engage with, interpret, and respond to this interaction stance. The study surfaces both opportunities -- such as vivid memory evocation, heightened sense of control, and creative exploration -- and critical tensions, including risks of false memory formation, algorithmic bias, and data privacy. Together, these findings articulate important boundaries for deploying generative AI in emotionally sensitive contexts.

Rapid progress in generative AI has given rise to Compound AI systems - pipelines comprised of multiple large language models (LLM), software tools and database systems. Compound AI systems are constructed on a layered traditional software stack running on a distributed hardware infrastructure. Many of the diverse software components are vulnerable to traditional security flaws documented in the Common Vulnerabilities and Exposures (CVE) database, while the underlying distributed hardware infrastructure remains exposed to timing attacks, bit-flip faults, and power-based side channels. Today, research targets LLM-specific risks like model extraction, training data leakage, and unsafe generation -- overlooking the impact of traditional system vulnerabilities. This work investigates how traditional software and hardware vulnerabilities can complement LLM-specific algorithmic attacks to compromise the integrity of a compound AI pipeline. We demonstrate two novel attacks that combine system-level vulnerabilities with algorithmic weaknesses: (1) Exploiting a software code injection flaw along with a guardrail Rowhammer attack to inject an unaltered jailbreak prompt into an LLM, resulting in an AI safety violation, and (2) Manipulating a knowledge database to redirect an LLM agent to transmit sensitive user data to a malicious application, thus breaching confidentiality. These attacks highlight the need to address traditional vulnerabilities; we systematize the attack primitives and analyze their composition by grouping vulnerabilities by their objective and mapping them to distinct stages of an attack lifecycle. This approach enables a rigorous red-teaming exercise and lays the groundwork for future defense strategies.

In decentralized Multi-Agent Reinforcement Learning (MARL), steganographic collusion -- where agents develop private protocols to evade monitoring -- presents a critical AI safety threat. Existing defenses, limited to behavioral or reward layers, fail to detect coordination in latent communication channels. We introduce the Dynamic Representational Circuit Breaker (DRCB), an architectural defense operating at the optimization substrate. Building on the AI Mother Tongue (AIM) framework, DRCB utilizes a Vector Quantized Variational Autoencoder (VQ-VAE) bottleneck to convert unobservable messages into auditable statistical objects. DRCB monitors signals including Jensen-Shannon Divergence drift, L2-norm codebook displacement, and Randomized Observer Pool accuracy to compute an EMA-based Collusion Score. Threshold breaches trigger four escalating interventions: dynamic adaptation, gradient-space penalty injection into the Advantage function A^pi, temporal reward suppression, and full substrate circuit breaking via codebook shuffling and optimizer state reset. Experiments on a Contextual Prisoner's Dilemma with MNIST labels show that while static monitoring fails (p = 0.3517), DRCB improves observer mean accuracy from 0.858 to 0.938 (+9.3 percent) and reduces volatility by 43 percent, while preserving mean joint reward (p = 0.854). Analysis of 214,298 symbol samples confirms "Semantic Degradation," where high-frequency sequences converge to zero entropy, foreclosing complex steganographic encodings. We identify a "Transparency Paradox" where agents achieve surface-level determinism while preserving residual capacity in long-tail distributions, reflecting Goodhart's Law. This task-agnostic methodology provides a technical path toward MICA-compliant (Multi-Agent Internal Coupling Audit) pre-deployment auditing for autonomous systems.

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.

Federated learning (FL) is an effective paradigm for distributed environments such as the Internet of Things (IoT), where data from diverse devices with varying functionalities remains localized while contributing to a shared global model. By eliminating the need to transmit raw data, FL inherently preserves privacy. However, the heterogeneous nature of IoT data, stemming from differences in device capabilities, data formats, and communication constraints, poses significant challenges to maintaining both global model performance and privacy. In the context of IoT-based anomaly detection, unsupervised FL offers a promising means to identify abnormal behavior without centralized data aggregation. Nevertheless, feature heterogeneity across devices complicates model training and optimization, hindering effective implementation. In this study we propose an efficient unsupervised FL framework that enhances anomaly detection by leveraging shared features from two distinct IoT datasets: one focused on anomaly detection and the other on device identification, while preserving dataset-specific features. To improve transparency and interpretability, we employ explainable AI techniques, such as SHAP, to identify key features influencing local model decisions. Experiments conducted on real-world IoT datasets demonstrate that the proposed method significantly outperforms conventional FL approaches in anomaly detection accuracy. This work underscores the potential of using shared features from complementary datasets to optimize unsupervised federated learning and achieve superior anomaly detection results in decentralized IoT environments.

While prior work has focused on projecting adversarial examples back onto the manifold of natural data to restore safety, we argue that a comprehensive understanding of AI safety requires characterizing the unsafe regions themselves. This paper introduces a framework for systematically mapping the Manifold of Failure in Large Language Models (LLMs). We reframe the search for vulnerabilities as a quality diversity problem, using MAP-Elites to illuminate the continuous topology of these failure regions, which we term behavioral attraction basins. Our quality metric, Alignment Deviation, guides the search towards areas where the model's behavior diverges most from its intended alignment. Across three LLMs: Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini, we show that MAP-Elites achieves up to 63% behavioral coverage, discovers up to 370 distinct vulnerability niches, and reveals dramatically different model-specific topological signatures: Llama-3-8B exhibits a near-universal vulnerability plateau (mean Alignment Deviation 0.93), GPT-OSS-20B shows a fragmented landscape with spatially concentrated basins (mean 0.73), and GPT-5-Mini demonstrates strong robustness with a ceiling at 0.50. Our approach produces interpretable, global maps of each model's safety landscape that no existing attack method (GCG, PAIR, or TAP) can provide, shifting the paradigm from finding discrete failures to understanding their underlying structure.

An Explainable Memory Forensics Approach for Malware Analysis

Silvia Lucia Sanna, Davide Maiorca, Giorgio Giacinto

Memory forensics is an effective methodology for analyzing living-off-the-land malware, including threats that employ evasion, obfuscation, anti-analysis, and steganographic techniques. By capturing volatile system state, memory analysis enables the recovery of transient artifacts such as decrypted payloads, executed commands, credentials, and cryptographic keys that are often inaccessible through static or traditional dynamic analysis. While several automated models have been proposed for malware detection from memory, their outputs typically lack interpretability, and memory analysis still relies heavily on expert-driven inspection of complex tool outputs, such as those produced by Volatility. In this paper, we propose an explainable, AI-assisted memory forensics approach that leverages general-purpose large language models (LLMs) to interpret memory analysis outputs in a human-readable form and to automatically extract meaningful Indicators of Compromise (IoCs), in some circumstances detecting more IoCs than current state-of-the-art tools. We apply the proposed methodology to both Windows and Android malware, comparing full RAM acquisition with target-process memory dumping and highlighting their complementary forensic value. Furthermore, we demonstrate how LLMs can support both expert and non-expert analysts by explaining analysis results, correlating artifacts, and justifying malware classifications. Finally, we show that a human-in-the-loop workflow, assisted by LLMs during kernel-assisted setup and analysis, improves reproducibility and reduces operational complexity, thereby reinforcing the practical applicability of AI-driven memory forensics for modern malware investigations.

Data privacy and eXplainable Artificial Intelligence (XAI) are two important aspects for modern Machine Learning systems. To enhance data privacy, recent machine learning models have been designed as a Federated Learning (FL) system. On top of that, additional privacy layers can be added, via Differential Privacy (DP). On the other hand, to improve explainability, ML must consider more interpretable approaches with reduced number of features and less complex internal architecture. In this context, this paper aims to achieve a machine learning (ML) model that combines enhanced data privacy with explainability. So, we propose a FL solution, called Federated EXplainable Trees with Differential Privacy (FEXT-DP), that: (i) is based on Decision Trees, since they are lightweight and have superior explainability than neural networks-based FL systems; (ii) provides additional layer of data privacy protection applying Differential Privacy (DP) to the Tree-Based model. However, there is a side effect adding DP: it harms the explainability of the system. So, this paper also presents the impact of DP protection on the explainability of the ML model, analyzing the obtained results for SHAP (SHapley Additive exPlanations) and Mean Decrease in Impurity (MDI).