Ofir Cohen

Daniel Kuznetsov, Ofir Cohen, Karin Shistik et al.

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.

Roie Kazoom, Ofir Cohen, Rami Puzis et al.

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.

Ofir Cohen, Gil Ari Agmon, Asaf Shabtai et al.

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 safety aspect of LLMs. ISA encompasses LLMs' security knowledge, which has been explored in the past, as well as attitudes and behaviors, which are crucial to LLMs' ability to understand implicit security context and reject unsafe requests that may cause the LLM to fail the user. We present an automated method for measuring the ISA of LLMs, which covers all 30 security topics in a mobile ISA taxonomy, using realistic scenarios that create tension between implicit security implications and user satisfaction. Applying this method to leading LLMs, we find that most of the popular models exhibit only medium to low levels of ISA, exposing their users to cybersecurity threats. Smaller variants of the same model family are significantly riskier, while newer versions show no consistent ISA improvement, suggesting that providers are not actively working toward mitigating this issue. These results reveal a widespread vulnerability affecting current LLM deployments: the majority of popular models, and particularly their smaller variants, may systematically endanger users. We propose a practical mitigation: incorporating our security awareness instruction into model system prompts to help LLMs better detect and reject unsafe requests.