Research Papers

Constraint solving is an elementary way for verification of deep neural networks (DNN). In the domain of AI safety, a DNN might be modified in its structure and parameters for its repair or attack. For such situations, we propose the incremental DNN verification problem, which asks whether a safety property still holds after the DNN is modified. To solve the problem, we present an incremental satisfiability modulo theory (SMT) algorithm based on the Reluplex framework. We simulate the most important features of the configurations that infers the verification result of the searching branches in the old solving procedure (with respect to the original network), and heuristically check whether the proofs are still valid for the modified DNN. We implement our algorithm as an incremental solver called DeepInc, and exerimental results show that DeepInc is more efficient in most cases. For the cases that the property holds both before and after modification, the acceleration can be faster by several orders of magnitude, showing that DeepInc is outstanding in incrementally searching for counterexamples. Moreover, based on the framework, we propose the multi-objective DNN repair problem and give an algorithm based on our incremental SMT solving algorithm. Our repair method preserves more potential safety properties on the repaired DNNs compared with state-of-the-art.

Sequential Nature of Recommender Systems Disrupts the Evaluation Process

Ali Shirali

Datasets are often generated in a sequential manner, where the previous samples and intermediate decisions or interventions affect subsequent samples. This is especially prominent in cases where there are significant human-AI interactions, such as in recommender systems. To characterize the importance of this relationship across samples, we propose to use adversarial attacks on popular evaluation processes. We present sequence-aware boosting attacks and provide a lower bound on the amount of extra information that can be exploited from a confidential test set solely based on the order of the observed data. We use real and synthetic data to test our methods and show that the evaluation process on the MovieLense-100k dataset can be affected by $\sim1\%$ which is important when considering the close competition. Codes are publicly available.

Ransomware detection systems increasingly rely on behavior-based machine learning to address evolving attack strategies. However, emerging privacy compliance, data governance, and responsible AI deployment demand not only accurate detection but also the ability to efficiently remove the influence of specific training samples without retraining the models from scratch. In this study, we present a privacy-aware machine unlearning evaluation framework for reinforcement learning (RL)-based ransomware detection built on Sharded, Isolated, Sliced, and Aggregated (SISA) training. The framework enables efficient data deletion by retraining only the affected model shards rather than the entire detector, reducing the retraining cost while preserving detection performance. We conduct a controlled comparative study using value-based RL agents, including Deep Q-Network (DQN) and Double Deep Q-Network (DDQN), under identical experimental settings with a cost-sensitive reward design and 5-fold cross-validation on Windows 11 ransomware dataset. Detection confidence is evaluated using a continuous Q-score margin, enabling ROC-AUC analysis beyond binary predictions. For unlearning, the dataset is partitioned into five shards with majority-vote aggregation, and a fast-unlearning path is evaluated by deleting 5% of the samples from a single shard and retraining only that shard. Results show that SISA-based unlearning incurs negligible utility degradation (<= 0.05 percent F1 drop) while substantially reducing retraining time relative to full SISA retraining. DDQN exhibits slightly improved stability and lower utility loss than DQN, while both agents maintain near identical in-distribution performance after unlearning. These findings indicate that SISA provides an efficient unlearning mechanism for RL-based ransomware detection, supporting privacy-aware deployment without compromising security effectiveness.

Cognitive radio networks (CRNs) are a key mechanism for alleviating spectrum scarcity by enabling secondary users (SUs) to opportunistically access licensed frequency bands without harmful interference to primary users (PUs). To address unreliable direct SU links and energy constraints common in next-generation wireless networks, this work introduces an adaptive, energy-aware hybrid reconfigurable intelligent surface (RIS) for underlay multiple-input single-output (MISO) CRNs. Distinct from prior approaches relying on static RIS architectures, our proposed RIS dynamically alternates between passive and active operation modes in real time according to harvested energy availability. We also model our scenario under practical hardware impairments and cascaded fading channels. We formulate and solve a joint transmit beamforming and RIS phase optimization problem via the soft actor-critic (SAC) deep reinforcement learning (DRL) method, leveraging its robustness in continuous and highly dynamic environments. Notably, we conduct the first systematic study of reward poisoning attacks on DRL agents in RIS-enhanced CRNs, and propose a lightweight, real-time defense based on reward clipping and statistical anomaly filtering. Numerical results demonstrate that the SAC-based approach consistently outperforms established DRL baselines, and that the dynamic hybrid RIS strikes a superior trade-off between throughput and energy consumption compared to fully passive and fully active alternatives. We further show the effectiveness of our defense in maintaining SU performance even under adversarial conditions. Our results advance the practical and secure deployment of RIS-assisted CRNs, and highlight crucial design insights for energy-constrained wireless systems.

Current frontier AI safety evaluations emphasize static benchmarks, third-party annotations, and red-teaming. In this position paper, we argue that AI safety research should focus on human-centered evaluations that measure harmful capability uplift: the marginal increase in a user's ability to cause harm with a frontier model beyond what conventional tools already enable. We frame harmful capability uplift as a core AI safety metric, ground it in prior social science research, and provide concrete methodological guidance for systematic measurement. We conclude with actionable steps for developers, researchers, funders, and regulators to make harmful capability uplift evaluation a standard practice.

Instruction tuned reasoning models are increasingly deployed with safety classifiers trained on frozen embeddings, assuming representation stability across model updates. We systematically investigate this assumption and find it fails: normalized perturbations of magnitude $σ=0.02$ (corresponding to $\approx 1^\circ$ angular drift on the embedding sphere) reduce classifier performance from $85\%$ to $50\%$ ROC-AUC. Critically, mean confidence only drops $14\%$, producing dangerous silent failures where $72\%$ of misclassifications occur with high confidence, defeating standard monitoring. We further show that instruction-tuned models exhibit 20$\%$ worse class separability than base models, making aligned systems paradoxically harder to safeguard. Our findings expose a fundamental fragility in production AI safety architectures and challenge the assumption that safety mechanisms transfer across model versions.

Intent Laundering: AI Safety Datasets Are Not What They Seem

Shahriar Golchin, Marc Wetter

We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three defining properties: being driven by ulterior intent, well-crafted, and out-of-distribution. We find that these datasets overrely on "triggering cues": words or phrases with overt negative/sensitive connotations that are intended to trigger safety mechanisms explicitly, which is unrealistic compared to real-world attacks. In practice, we evaluate whether these datasets genuinely measure safety risks or merely provoke refusals through triggering cues. To explore this, we introduce "intent laundering": a procedure that abstracts away triggering cues from adversarial attacks (data points) while strictly preserving their malicious intent and all relevant details. Our results show that current adversarial safety datasets fail to faithfully represent real-world adversarial behavior due to their overreliance on triggering cues. Once these cues are removed, all previously evaluated "reasonably safe" models become unsafe, including Gemini 3 Pro and Claude Sonnet 3.7/4. Moreover, when intent laundering is adapted as a jailbreaking technique, it consistently achieves high attack success rates, ranging from 90.00% to 100.00%, under fully black-box access. Overall, our findings expose a significant disconnect between how existing datasets evaluate model safety and how real-world adversaries behave.

Current agentic AI architectures are fundamentally incompatible with the security and epistemological requirements of high-stakes scientific workflows. The problem is not inadequate alignment or insufficient guardrails, it is architectural: autoregressive language models process all tokens uniformly, making deterministic command--data separation unattainable through training alone. We argue that deterministic, architectural enforcement, not probabilistic learned behavior, is a necessary condition for trustworthy AI-assisted science. We introduce the Trinity Defense Architecture, which enforces security through three mechanisms: action governance via a finite action calculus with reference-monitor enforcement, information-flow control via mandatory access labels preventing cross-scope leakage, and privilege separation isolating perception from execution. We show that without unforgeable provenance and deterministic mediation, the ``Lethal Trifecta'' (untrusted inputs, privileged data access, external action capability) turns authorization security into an exploit-discovery problem: training-based defenses may reduce empirical attack rates but cannot provide deterministic guarantees. The ML community must recognize that alignment is insufficient for authorization security, and that architectural mediation is required before agentic AI can be safely deployed in consequential scientific domains.

Open-weight language models are increasingly used in production settings, raising new security challenges. One prominent threat is backdoor attacks, in which adversaries embed hidden behaviors that activate under specific conditions. Previous work has assumed that adversaries have access to training pipelines or deployment infrastructure. We propose a novel attack surface requiring neither: the "chat template". Chat templates are executable programs invoked at every inference call, often implemented in Jinja2, that occupy a privileged position between user input and model processing. We show that an adversary who distributes a model with a maliciously modified template can implant an inference-time backdoor without modifying model weights, poisoning training data, or controlling runtime infrastructure. We evaluate this attack across three deployment tiers. At the LLM level, triggered backdoors reduce factual accuracy from 90% to 15% on average and induce attacker-controlled URL emission with success rates exceeding 80%, while benign inputs show no measurable degradation; these results hold across eighteen models. At the agent level, template backdoors hijack tool-use across two benchmarks spanning 3,868 episodes, bypassing every tested injection defense offered by the benchmarks while remaining fully dormant absent the trigger. At the multi-agent system level, we demonstrate how a single poisoned artifact compromises a real-world agentic deployment and propagates supply-chain code poisoning downstream. The poisoned artifacts evade all security scans on the largest open model distribution platform; and because the payload is rendered by the template before user input is processed, it is architecturally unreachable by input-level defenses such as prompt injection guardrails. These results establish chat templates as a reliable and undefended attack in the open-weight AI supply chain.

Large language models (LLMs) have made it remarkably easy to synthesize plausible source code from natural language prompts. While this accelerates software development and supports learning, it also raises new risks for academic integrity, authorship attribution, and responsible AI use. This paper investigates the problem of distinguishing human-written from machine-generated code by comparing two complementary approaches: feature-based detectors built from lightweight, interpretable stylometric and structural properties of code, and embedding-based detectors leveraging pretrained code encoders. Using a recent large-scale benchmark dataset of 600k human-written and AI-generated code samples, we find that feature-based models achieve strong performance (ROC-AUC 0.995, PR-AUC 0.995, F1 0.971), while embedding-based models with CodeBERT embeddings are also very competitive (ROC-AUC 0.994, PR-AUC 0.994, F1 0.965). Analysis shows that features tied to indentation and whitespace provide particularly discriminative cues, whereas embeddings capture deeper semantic patterns and yield slightly higher precision. These findings underscore the trade-offs between interpretability and generalization, offering practical guidance for deploying robust code-origin detection in academic and industrial contexts.

Brain foundation models bring the foundation model paradigm to the field of neuroscience. Like language and image foundation models, they are general-purpose AI systems pretrained on large-scale datasets that adapt readily to downstream tasks. Unlike text-and-image based models, however, they train on brain data: large-datasets of EEG, fMRI, and other neural data types historically collected within tightly governed clinical and research settings. This paper contends that training foundation models on neural data opens new normative territory. Neural data carry stronger expectations of, and claims to, protection than text or images, given their body-derived nature and historical governance within clinical and research settings. Yet the foundation model paradigm subjects them to practices of large-scale repurposing, cross-context stitching, and open-ended downstream application. Furthermore, these practices are now accessible to a much broader range of actors, including commercial developers, against a backdrop of fragmented and unclear governance. To map this territory, we first describe brain foundation models' technical foundations and training-data ecosystem. We then draw on AI ethics, neuroethics, and bioethics to organize concerns across privacy, consent, bias, benefit sharing, and governance. For each, we propose both agenda-setting questions and baseline safeguards as the field matures.

While recent automated red-teaming methods show promise for systematically exposing model vulnerabilities, most existing approaches rely on human-specified workflows. This dependence on manually designed workflows suffers from human biases and makes exploring the broader design space expensive. We introduce AgenticRed, an automated pipeline that leverages LLMs' in-context learning to iteratively design and refine red-teaming systems without human intervention. Rather than optimizing attacker policies within predefined structures, AgenticRed treats red-teaming as a system design problem, and it autonomously evolves automated red-teaming systems using evolutionary selection and generational knowledge. Red-teaming systems designed by AgenticRed consistently outperform state-of-the-art approaches, achieving 96% attack success rate (ASR) on Llama-2-7B, 98% on Llama-3-8B and 100% on Qwen3-8B on HarmBench. Our approach generates robust, query-agnostic red-teaming systems that transfer strongly to the latest proprietary models, achieving an impressive 100% ASR on GPT-5.1, DeepSeek-R1 and DeepSeek V3.2. This work highlights evolutionary algorithms as a powerful approach to AI safety that can keep pace with rapidly evolving models.