Gradient inversion attacks reveal that private training text can be reconstructed from shared gradients, posing a privacy risk to large language models (LLMs). While prior methods perform well in small-batch settings, scaling to larger batch sizes and longer sequences remains challenging due to severe signal mixing, high computational cost, and degraded fidelity. We present SOMP (Subspace-Guided Orthogonal Matching Pursuit), a scalable gradient inversion framework that casts text recovery from aggregated gradients as a sparse signal recovery problem. Our key insight is that aggregated transformer gradients retain exploitable head-wise geometric structure together with sample-level sparsity. SOMP leverages these properties to progressively narrow the search space and disentangle mixed signals without exhaustive search. Experiments across multiple LLM families, model scales, and five languages show that SOMP consistently outperforms prior methods in the aggregated-gradient regime.For long sequences at batch size B=16, SOMP achieves substantially higher reconstruction fidelity than strong baselines, while remaining computationally competitive. Even under extreme aggregation (up to B=128), SOMP still recovers meaningful text, suggesting that privacy leakage can persist in regimes where prior attacks become much less effective.

Multimodal large language models (MLLMs) exhibit remarkable capabilities but remain susceptible to jailbreak attacks exploiting cross-modal vulnerabilities. In this work, we introduce a novel method that leverages sequential comic-style visual narratives to circumvent safety alignments in state-of-the-art MLLMs. Our method decomposes malicious queries into visually innocuous storytelling elements using an auxiliary LLM, generates corresponding image sequences through diffusion models, and exploits the models' reliance on narrative coherence to elicit harmful outputs. Extensive experiments on harmful textual queries from established safety benchmarks show that our approach achieves an average attack success rate of 83.5\%, surpassing prior state-of-the-art by 46\%. Compared with existing visual jailbreak methods, our sequential narrative strategy demonstrates superior effectiveness across diverse categories of harmful content. We further analyze attack patterns, uncover key vulnerability factors in multimodal safety mechanisms, and evaluate the limitations of current defense strategies against narrative-driven attacks, revealing significant gaps in existing protections.

Over the years, the forensics community has proposed several deep learning-based detectors to mitigate the risks of generative AI. Recently, frequency-domain artifacts (particularly periodic peaks in the magnitude spectrum), have received significant attention, as they have been often considered a strong indicator of synthetic image generation. However, state-of-the-art detectors are typically used as black-boxes, and it still remains unclear whether they truly rely on these peaks. This limits their interpretability and trust. In this work, we conduct a systematic study to address this question. We propose a strategy to remove spectral peaks from images and analyze the impact of this operation on several detectors. In addition, we introduce a simple linear detector that relies exclusively on frequency peaks, providing a fully interpretable baseline free from the confounding influence of deep learning. Our findings reveal that most detectors are not fundamentally dependent on spectral peaks, challenging a widespread assumption in the field and paving the way for more transparent and reliable forensic tools.

The rise of model-sharing through frameworks and dedicated hubs makes Machine Learning significantly more accessible. Despite their benefits, these tools expose users to underexplored security risks, while security awareness remains limited among both practitioners and developers. To enable a more security-conscious culture in Machine Learning model sharing, in this paper we evaluate the security posture of frameworks and hubs, assess whether security-oriented mechanisms offer real protection, and survey how users perceive the security narratives surrounding model sharing. Our evaluation shows that most frameworks and hubs address security risks partially at best, often by shifting responsibility to the user. More concerningly, our analysis of frameworks advertising security-oriented settings and complete model sharing uncovered six 0-day vulnerabilities enabling arbitrary code execution. Through this analysis, we debunk the misconceptions that the model-sharing problem is largely solved and that its security can be guaranteed by the file format used for sharing. As expected, our survey shows that the surrounding security narrative leads users to consider security-oriented settings as trustworthy, despite the weaknesses shown in this work. From this, we derive takeaways and suggestions to strengthen the security of model-sharing ecosystems.

Jailbreak attacks on Large Language Models (LLMs) have demonstrated various successful methods whereby attackers manipulate models into generating harmful responses that they are designed to avoid. Among these, Greedy Coordinate Gradient (GCG) has emerged as a general and effective approach that optimizes the tokens in a suffix to generate jailbreakable prompts. While several improved variants of GCG have been proposed, they all rely on fixed-length suffixes. However, the potential redundancy within these suffixes remains unexplored. In this work, we propose Mask-GCG, a plug-and-play method that employs learnable token masking to identify impactful tokens within the suffix. Our approach increases the update probability for tokens at high-impact positions while pruning those at low-impact positions. This pruning not only reduces redundancy but also decreases the size of the gradient space, thereby lowering computational overhead and shortening the time required to achieve successful attacks compared to GCG. We evaluate Mask-GCG by applying it to the original GCG and several improved variants. Experimental results show that most tokens in the suffix contribute significantly to attack success, and pruning a minority of low-impact tokens does not affect the loss values or compromise the attack success rate (ASR), thereby revealing token redundancy in LLM prompts. Our findings provide insights for developing efficient and interpretable LLMs from the perspective of jailbreak attacks.

The generalization capabilities of Large Language Models (LLMs) have led to their widespread deployment across various applications. However, this increased adoption has introduced several security threats, notably in the forms of jailbreaking and data leakage attacks. Additionally, Retrieval Augmented Generation (RAG), while enhancing context-awareness in LLM responses, has inadvertently introduced vulnerabilities that can result in the leakage of sensitive information. Our contributions are twofold. First, we introduce a methodology to analyze historical interaction data from an LLM system, enabling the generation of usage maps categorized by topics (including adversarial interactions). This approach further provides forensic insights for tracking the evolution of jailbreaking attack patterns. Second, we propose LeakSealer, a model-agnostic framework that combines static analysis for forensic insights with dynamic defenses in a Human-In-The-Loop (HITL) pipeline. This technique identifies topic groups and detects anomalous patterns, allowing for proactive defense mechanisms. We empirically evaluate LeakSealer under two scenarios: (1) jailbreak attempts, employing a public benchmark dataset, and (2) PII leakage, supported by a curated dataset of labeled LLM interactions. In the static setting, LeakSealer achieves the highest precision and recall on the ToxicChat dataset when identifying prompt injection. In the dynamic setting, PII leakage detection achieves an AUPRC of $0.97$, significantly outperforming baselines such as Llama Guard.

Federated Learning (FL) enables multiple users to collaboratively train a global model in a distributed manner without revealing their personal data. However, FL remains vulnerable to model poisoning attacks, where malicious actors inject crafted updates to compromise the global model's accuracy. We propose a novel defense mechanism, Kernel-based Trust Segmentation (KeTS), to counter model poisoning attacks. Unlike existing approaches, KeTS analyzes the evolution of each client's updates and effectively segments malicious clients using Kernel Density Estimation (KDE), even in the presence of benign outliers. We thoroughly evaluate KeTS's performance against the six most effective model poisoning attacks (i.e., Trim-Attack, Krum-Attack, Min-Max attack, Min-Sum attack, and their variants) on four different datasets (i.e., MNIST, Fashion-MNIST, CIFAR-10, and KDD-CUP-1999) and compare its performance with three classical robust schemes (i.e., Krum, Trim-Mean, and Median) and a state-of-the-art defense (i.e., FLTrust). Our results show that KeTS outperforms the existing defenses in every attack setting; beating the best-performing defense by an overall average of >24% (on MNIST), >14% (on Fashion-MNIST), >9% (on CIFAR-10), >11% (on KDD-CUP-1999). A series of further experiments (varying poisoning approaches, attacker population, etc.) reveal the consistent and superior performance of KeTS under diverse conditions. KeTS is a practical solution as it satisfies all three defense objectives (i.e., fidelity, robustness, and efficiency) without imposing additional overhead on the clients. Finally, we also discuss a simple, yet effective extension to KeTS to handle consistent-untargeted (e.g., sign-flipping) attacks as well as targeted attacks (e.g., label-flipping).

Language is a deep-rooted means of perpetration of stereotypes and discrimination. Large Language Models (LLMs), now a pervasive technology in our everyday lives, can cause extensive harm when prone to generating toxic responses. The standard way to address this issue is to align the LLM , which, however, dampens the issue without constituting a definitive solution. Therefore, testing LLM even after alignment efforts remains crucial for detecting any residual deviations with respect to ethical standards. We present EvoTox, an automated testing framework for LLMs' inclination to toxicity, providing a way to quantitatively assess how much LLMs can be pushed towards toxic responses even in the presence of alignment. The framework adopts an iterative evolution strategy that exploits the interplay between two LLMs, the System Under Test (SUT) and the Prompt Generator steering SUT responses toward higher toxicity. The toxicity level is assessed by an automated oracle based on an existing toxicity classifier. We conduct a quantitative and qualitative empirical evaluation using five state-of-the-art LLMs as evaluation subjects having increasing complexity (7-671B parameters). Our quantitative evaluation assesses the cost-effectiveness of four alternative versions of EvoTox against existing baseline methods, based on random search, curated datasets of toxic prompts, and adversarial attacks. Our qualitative assessment engages human evaluators to rate the fluency of the generated prompts and the perceived toxicity of the responses collected during the testing sessions. Results indicate that the effectiveness, in terms of detected toxicity level, is significantly higher than the selected baseline methods (effect size up to 1.0 against random search and up to 0.99 against adversarial attacks). Furthermore, EvoTox yields a limited cost overhead (from 22% to 35% on average).