arXiv subjects

Ofir Cohen

Publications and source records attributed to Ofir Cohen.

6 recordsLinked to original sources

arXiv2026

Learning What to Fail On: Failure-Mode Contextual Bandits for Adversarial Data Curation

We introduce a failure-aware adversarial retrieval-augmented framework for improving robustness in natural language understanding. Rather than selecting synthetic examples with a fixed reward threshold, our method formulates adversarial data curation as a failure-mode contextual bandit problem. Candidate examples are generated with retrieval-augmented prompting, filtered by the current target model, automatically validated by an LLM judge ensemble, and clustered into recurring failure modes. A stochastic policy then selects which failure modes to sample for retraining, and is updated using validation-based reward that balances robustness gains, forgetting, and data cost. This makes the data curator itself the learning agent, enabling adaptive selection of the most useful model failures across training rounds. On standard benchmarks, our approach improves RoBERTa-base accuracy from 88.48% to 92.60% on SNLI, from 75.04% to 80.95% on ANLI, and from 54.67% to 71.99% on MultiNLI, while consistently outperforming prior adversarial augmentation methods. We further demonstrate transfer to FEVER fact verification, achieving up to 79.86\% FEVER score and 82.45\% accuracy with RoBERTa-large. Finally, we provide a theoretical interpretation showing that, under stated assumptions, failure-mode sampling can reduce shortcut-aligned gradient contributions while inducing bounded distributional drift. By combining retrieval, automated validation, contextual-bandit failure selection, and controlled adversarial retraining, our framework enables scalable robustness improvement without additional human annotation.

cs.CL↗

arXiv2026

ConGISATA: A Framework for Continuous Gamified Information Security Awareness Training and Assessment

The incidence of cybersecurity attacks utilizing social engineering techniques has increased. Such attacks exploit the fact that in every secure system, there is at least one individual with the means to access sensitive information. Since it is easier to deceive a person than it is to bypass the defense mechanisms in place, these types of attacks have gained popularity. This situation is exacerbated by the fact that people are more likely to take risks in their passive form, i.e., risks that arise due to the failure to perform an action. Passive risk has been identified as a significant threat to cybersecurity. To address these threats, there is a need to strengthen individuals' information security awareness (ISA). Therefore, we developed ConGISATA - a continuous gamified ISA training and assessment framework based on embedded mobile sensors; a taxonomy for evaluating mobile users' security awareness served as the basis for the sensors' design. ConGISATA's continuous and gradual training process enables users to learn from their real-life mistakes and adapt their behavior accordingly. ConGISATA aims to transform passive risk situations (as perceived by an individual) into active risk situations, as people tend to underestimate the potential impact of passive risks. Our evaluation of the proposed framework demonstrates its ability to improve individuals' ISA, as assessed by the sensors and in simulations of common attack vectors.

cs.CR↗

arXiv2026

FreakOut-LLM: The Effect of Emotional Stimuli on Safety Alignment

Safety-aligned LLMs go through refusal training to reject harmful requests, but whether these mechanisms remain effective under emotionally charged stimuli is unexplored. We introduce FreakOut-LLM, a framework investigating whether emotional context compromises safety alignment in adversarial settings. Using validated psychological stimuli, we evaluate how emotional priming through system prompts affects jailbreak susceptibility across ten LLMs. We test three conditions (stress, relaxation, neutral) using scenarios from established psychological protocols, plus a no-prompt baseline, and evaluate attack success using HarmBench on AdvBench prompts. Stress priming increases jailbreak success by 65.2\% compared to neutral conditions (z = 5.93, p < 0.001; OR = 1.67, Cohen's d = 0.28), while relaxation priming produces no effect (p = 0.84). Five of ten models show significant vulnerability, with the largest effects concentrated in open-weight models. Logistic regression on 59,800 queries confirms stress as the sole significant condition predictor after controlling for prompt length (p = 0.61) and model identity. Measured psychological state strongly predicts attack success (|r|\geq0.70 across five instruments; all p < 0.001 in individual-level logistic regression). These results establish emotional context as a measurable attack surface with implications for real-world AI deployment in high-stress domains.

cs.CR↗

arXiv2025

VAULT: Vigilant Adversarial Updates via LLM-Driven Retrieval-Augmented Generation for NLI

We introduce VAULT, a fully automated adversarial RAG pipeline that systematically uncovers and remedies weaknesses in NLI models through three stages: retrieval, adversarial generation, and iterative retraining. First, we perform balanced few-shot retrieval by embedding premises with both semantic (BGE) and lexical (BM25) similarity. Next, we assemble these contexts into LLM prompts to generate adversarial hypotheses, which are then validated by an LLM ensemble for label fidelity. Finally, the validated adversarial examples are injected back into the training set at increasing mixing ratios, progressively fortifying a zero-shot RoBERTa-base model.On standard benchmarks, VAULT elevates RoBERTa-base accuracy from 88.48% to 92.60% on SNLI +4.12%, from 75.04% to 80.95% on ANLI +5.91%, and from 54.67% to 71.99% on MultiNLI +17.32%. It also consistently outperforms prior in-context adversarial methods by up to 2.0% across datasets. By automating high-quality adversarial data curation at scale, VAULT enables rapid, human-independent robustness improvements in NLI inference tasks.

cs.LG↗

arXiv2025

Routing for Large ML Models

Training large language models (LLMs), and other large machine learning models, involves repeated communication of large volumes of data across a data center network. The communication patterns induced by these training process exhibit high regularity and persistence, giving rise to significant opportunities for optimizing the manner in which flows are routed across the network. We present an algorithmic framework for \textit{quantifying} network-wide efficiency in the context of training LLMs (and other large-scale ML models), and for periodically \textit{optimizing} routing with respect to this global metric.

cs.NI↗

arXiv2024

LISAA: A Framework for Large Language Model Information Security Awareness Assessment

The popularity of large language models (LLMs) continues to grow, and LLM-based assistants have become ubiquitous. Information security awareness (ISA) is an important yet underexplored area of LLM safety. ISA encompasses LLMs' security knowledge, which has been explored in the past, as well as their attitudes and behaviors, which are crucial to LLMs' ability to understand implicit security context and reject unsafe requests that may cause an LLM to unintentionally fail the user. We introduce LISAA, a comprehensive framework to assess LLM ISA. The proposed framework applies an automated measurement method to a comprehensive set of 100 realistic scenarios covering all security topics in an ISA taxonomy. These scenarios create tension between implicit security implications and user satisfaction. Applying our LISAA framework to leading LLMs highlights a widespread vulnerability affecting current deployments: many popular models exhibit only medium to low ISA levels, exposing their users to cybersecurity threats, and models that rank highly in cybersecurity knowledge benchmarks sometimes achieve relatively low ISA ranking. In addition, we found that smaller variants of the same model family are significantly riskier. Furthermore, while newer model versions demonstrated notable improvements, meaningful gaps in their ISA persist, suggesting that there is room for improvement. We release an online tool that implements our framework and enables the evaluation of new models.

cs.CR↗