Voice anonymization masks vocal traits while preserving linguistic content, which may still leak speaker-specific patterns. To assess and strengthen privacy evaluation, we propose a dual-stream attacker that fuses spectral and self-supervised learning features via parallel encoders with a three-stage training strategy. Stage I establishes foundational speaker-discriminative representations. Stage II leverages the shared identity-transformation characteristics of voice conversion and anonymization, exposing the model to diverse converted speech to build cross-system robustness. Stage III provides lightweight adaptation to target anonymized data. Results on the VoicePrivacy Attacker Challenge (VPAC) dataset demonstrate that Stage II is the primary driver of generalization, enabling strong attacking performance on unseen anonymization datasets. With Stage III, fine-tuning on only 10\% of the target anonymization dataset surpasses current state-of-the-art attackers in terms of EER.

Training data leakage from Large Language Models (LLMs) raises serious concerns related to privacy, security, and copyright compliance. A central challenge in assessing this risk is distinguishing genuine memorization of training data from the generation of statistically common sequences. Existing approaches to measuring memorization often conflate these phenomena, labeling outputs as memorized even when they arise from generalization over common patterns. Counterfactual Memorization provides a principled solution by comparing models trained with and without a target sequence, but its reliance on retraining multiple baseline models makes it computationally expensive and impractical at scale. This work introduces Prior-Aware Memorization, a theoretically grounded, lightweight and training-free criterion for identifying genuine memorization in LLMs. The key idea is to evaluate whether a candidate suffix is strongly associated with its specific training prefix or whether it appears with high probability across many unrelated prompts due to statistical commonality. We evaluate this metric on text from the training corpora of two pre-trained models, LLaMA and OPT, using both long sequences (to simulate copyright risks) and named entities (to simulate PII leakage). Our results show that between 55% and 90% of sequences previously labeled as memorized are in fact statistically common. Similar findings hold for the SATML training data extraction challenge dataset, where roughly 40% of sequences exhibit common-pattern behavior despite appearing only once in the training data. These results demonstrate that low frequency alone is insufficient evidence of memorization and highlight the importance of accounting for model priors when assessing leakage.

% Prompt injection attacks insert malicious instructions into an LLM's input to steer it toward an attacker-chosen task instead of the intended one. Existing detection defenses typically classify any input with instruction as malicious, leading to misclassification of benign inputs containing instructions that align with the intended task. In this work, we account for the instruction hierarchy and distinguish among three categories: inputs with misaligned instructions, inputs with aligned instructions, and non-instruction inputs. We introduce AlignSentinel, a three-class classifier that leverages features derived from LLM's attention maps to categorize inputs accordingly. To support evaluation, we construct the first systematic benchmark containing inputs from all three categories. Experiments on both our benchmark and existing ones--where inputs with aligned instructions are largely absent--show that AlignSentinel accurately detects inputs with misaligned instructions and substantially outperforms baselines.

The evolution of Large Language Models (LLMs) has resulted in a paradigm shift towards autonomous agents, necessitating robust security against Prompt Injection (PI) vulnerabilities where untrusted inputs hijack agent behaviors. This SoK presents a comprehensive overview of the PI landscape, covering attacks, defenses, and their evaluation practices. Through a systematic literature review and quantitative analysis, we establish taxonomies that categorize PI attacks by payload generation strategies (heuristic vs. optimization) and defenses by intervention stages (text, model, and execution levels). Our analysis reveals a key limitation shared by many existing defenses and benchmarks: they largely overlook context-dependent tasks, in which agents are authorized to rely on runtime environmental observations to determine actions. To address this gap, we introduce AgentPI, a new benchmark designed to systematically evaluate agent behavior under context-dependent interaction settings. Using AgentPI, we empirically evaluate representative defenses and show that no single approach can simultaneously achieve high trustworthiness, high utility, and low latency. Moreover, we show that many defenses appear effective under existing benchmarks by suppressing contextual inputs, yet fail to generalize to realistic agent settings where context-dependent reasoning is essential. This SoK distills key takeaways and open research problems, offering structured guidance for future research and practical deployment of secure LLM agents.

Large reasoning models (LRMs) extend large language models with explicit multi-step reasoning traces, but this capability introduces a new class of prompt-induced inference-time denial-of-service (PI-DoS) attacks that exploit the high computational cost of reasoning. We first formalize inference cost for LRMs and define PI-DoS, then prove that any practical PI-DoS attack should satisfy three properties: (1) a high amplification ratio, where each query induces a disproportionately long reasoning trace relative to its own length; (ii) stealthiness, in which prompts and responses remain on the natural language manifold and evade distribution shift detectors; and (iii) optimizability, in which the attack supports efficient optimization without being slowed by its own success. Under this framework, we present ReasoningBomb, a reinforcement-learning-based PI-DoS framework that is guided by a constant-time surrogate reward and trains a large reasoning-model attacker to generate short natural prompts that drive victim LRMs into pathologically long and often effectively non-terminating reasoning. Across seven open-source models (including LLMs and LRMs) and three commercial LRMs, ReasoningBomb induces 18,759 completion tokens on average and 19,263 reasoning tokens on average across reasoning models. It outperforms the the runner-up baseline by 35% in completion tokens and 38% in reasoning tokens, while inducing 6-7x more tokens than benign queries and achieving 286.7x input-to-output amplification ratio averaged across all samples. Additionally, our method achieves 99.8% bypass rate on input-based detection, 98.7% on output-based detection, and 98.4% against strict dual-stage joint detection.

Watermarking (WM) is a critical mechanism for detecting and attributing AI-generated content. Current WM methods for Large Language Models (LLMs) are predominantly tailored for autoregressive (AR) models: They rely on tokens being generated sequentially, and embed stable signals within the generated sequence based on the previously sampled text. Diffusion Language Models (DLMs) generate text via non-sequential iterative denoising, which requires significant modification to use WM methods designed for AR models. Recent work proposed to watermark DLMs by inverting the process when needed, but suffers significant computational or memory overhead. We introduce Left-Right Diffusion Watermarking (LR-DWM), a scheme that biases the generated token based on both left and right neighbors, when they are available. LR-DWM incurs minimal runtime and memory overhead, remaining close to the non-watermarked baseline DLM while enabling reliable statistical detection under standard evaluation settings. Our results demonstrate that DLMs can be watermarked efficiently, achieving high detectability with negligible computational and memory overhead.

Large Language Models (LLMs) have enabled the development of powerful agentic systems capable of automating complex workflows across various fields. However, these systems are highly vulnerable to indirect prompt injection attacks, where malicious instructions embedded in external data can hijack agent behavior. In this work, we present ReasAlign, a model-level solution to improve safety alignment against indirect prompt injection attacks. The core idea of ReasAlign is to incorporate structured reasoning steps to analyze user queries, detect conflicting instructions, and preserve the continuity of the user's intended tasks to defend against indirect injection attacks. To further ensure reasoning logic and accuracy, we introduce a test-time scaling mechanism with a preference-optimized judge model that scores reasoning steps and selects the best trajectory. Comprehensive evaluations across various benchmarks show that ReasAlign maintains utility comparable to an undefended model while consistently outperforming Meta SecAlign, the strongest prior guardrail. On the representative open-ended CyberSecEval2 benchmark, which includes multiple prompt-injected tasks, ReasAlign achieves 94.6% utility and only 3.6% ASR, far surpassing the state-of-the-art defensive model of Meta SecAlign (56.4% utility and 74.4% ASR). These results demonstrate that ReasAlign achieves the best trade-off between security and utility, establishing a robust and practical defense against prompt injection attacks in real-world agentic systems. Our code and experimental results could be found at https://github.com/leolee99/ReasAlign.

Indirect prompt injection attacks (IPIAs), where large language models (LLMs) follow malicious instructions hidden in input data, pose a critical threat to LLM-powered agents. In this paper, we present IntentGuard, a general defense framework based on instruction-following intent analysis. The key insight of IntentGuard is that the decisive factor in IPIAs is not the presence of malicious text, but whether the LLM intends to follow instructions from untrusted data. Building on this insight, IntentGuard leverages an instruction-following intent analyzer (IIA) to identify which parts of the input prompt the model recognizes as actionable instructions, and then flag or neutralize any overlaps with untrusted data segments. To instantiate the framework, we develop an IIA that uses three "thinking intervention" strategies to elicit a structured list of intended instructions from reasoning-enabled LLMs. These techniques include start-of-thinking prefilling, end-of-thinking refinement, and adversarial in-context demonstration. We evaluate IntentGuard on two agentic benchmarks (AgentDojo and Mind2Web) using two reasoning-enabled LLMs (Qwen-3-32B and gpt-oss-20B). Results demonstrate that IntentGuard achieves (1) no utility degradation in all but one setting and (2) strong robustness against adaptive prompt injection attacks (e.g., reducing attack success rates from 100% to 8.5% in a Mind2Web scenario).

This paper introduces a dynamic and actionable framework for securing agentic AI systems in enterprise deployment. We contend that safety and security are not merely fixed attributes of individual models but also emergent properties arising from the dynamic interactions among models, orchestrators, tools, and data within their operating environments. We propose a new way of identification of novel agentic risks through the lens of user safety. Although, for traditional LLMs and agentic models in isolation, safety and security has a clear separation, through the lens of safety in agentic systems, they appear to be connected. Building on this foundation, we define an operational agentic risk taxonomy that unifies traditional safety and security concerns with novel, uniquely agentic risks, including tool misuse, cascading action chains, and unintended control amplification among others. At the core of our approach is a dynamic agentic safety and security framework that operationalizes contextual agentic risk management by using auxiliary AI models and agents, with human oversight, to assist in contextual risk discovery, evaluation, and mitigation. We further address one of the most challenging aspects of safety and security of agentic systems: risk discovery through sandboxed, AI-driven red teaming. We demonstrate the framework effectiveness through a detailed case study of NVIDIA flagship agentic research assistant, AI-Q Research Assistant, showcasing practical, end-to-end safety and security evaluations in complex, enterprise-grade agentic workflows. This risk discovery phase finds novel agentic risks that are then contextually mitigated. We also release the dataset from our case study, containing traces of over 10,000 realistic attack and defense executions of the agentic workflow to help advance research in agentic safety.

We present Contamination Detection via Context (CoDeC), a practical and accurate method to detect and quantify training data contamination in large language models. CoDeC distinguishes between data memorized during training and data outside the training distribution by measuring how in-context learning affects model performance. We find that in-context examples typically boost confidence for unseen datasets but may reduce it when the dataset was part of training, due to disrupted memorization patterns. Experiments show that CoDeC produces interpretable contamination scores that clearly separate seen and unseen datasets, and reveals strong evidence of memorization in open-weight models with undisclosed training corpora. The method is simple, automated, and both model- and dataset-agnostic, making it easy to integrate with benchmark evaluations.

Large audio-language models (LALMs) extend text-based LLMs with auditory understanding, offering new opportunities for multimodal applications. While their perception, reasoning, and task performance have been widely studied, their safety alignment under paralinguistic variation remains underexplored. This work systematically investigates the role of speaker emotion. We construct a dataset of malicious speech instructions expressed across multiple emotions and intensities, and evaluate several state-of-the-art LALMs. Our results reveal substantial safety inconsistencies: different emotions elicit varying levels of unsafe responses, and the effect of intensity is non-monotonic, with medium expressions often posing the greatest risk. These findings highlight an overlooked vulnerability in LALMs and call for alignment strategies explicitly designed to ensure robustness under emotional variation, a prerequisite for trustworthy deployment in real-world settings.

The widespread adoption of Retrieval-Augmented Image Generation (RAIG) has raised significant concerns about the unauthorized use of private image datasets. While these systems have shown remarkable capabilities in enhancing generation quality through reference images, protecting visual datasets from unauthorized use in such systems remains a challenging problem. Traditional digital watermarking approaches face limitations in RAIG systems, as the complex feature extraction and recombination processes fail to preserve watermark signals during generation. To address these challenges, we propose ImageSentinel, a novel framework for protecting visual datasets in RAIG. Our framework synthesizes sentinel images that maintain visual consistency with the original dataset. These sentinels enable protection verification through randomly generated character sequences that serve as retrieval keys. To ensure seamless integration, we leverage vision-language models to generate the sentinel images. Experimental results demonstrate that ImageSentinel effectively detects unauthorized dataset usage while preserving generation quality for authorized applications. Code is available at https://github.com/luo-ziyuan/ImageSentinel.

AI-based talking-head videoconferencing systems reduce bandwidth by sending a compact pose-expression latent and re-synthesizing RGB at the receiver, but this latent can be puppeteered, letting an attacker hijack a victim's likeness in real time. Because every frame is synthetic, deepfake and synthetic video detectors fail outright. To address this security problem, we exploit a key observation: the pose-expression latent inherently contains biometric information of the driving identity. Therefore, we introduce the first biometric leakage defense without ever looking at the reconstructed RGB video: a pose-conditioned, large-margin contrastive encoder that isolates persistent identity cues inside the transmitted latent while cancelling transient pose and expression. A simple cosine test on this disentangled embedding flags illicit identity swaps as the video is rendered. Our experiments on multiple talking-head generation models show that our method consistently outperforms existing puppeteering defenses, operates in real-time, and shows strong generalization to out-of-distribution scenarios.

Retrieval-Augmented Generation (RAG) enhances Large Language Models by grounding their outputs in external documents. These systems, however, remain vulnerable to attacks on the retrieval corpus, such as prompt injection. RAG-based search systems (e.g., Google's Search AI Overview) present an interesting setting for studying and protecting against such threats, as defense algorithms can benefit from built-in reliability signals -- like document ranking -- and represent a non-LLM challenge for the adversary due to decades of work to thwart SEO. Motivated by, but not limited to, this scenario, this work introduces ReliabilityRAG, a framework for adversarial robustness that explicitly leverages reliability information of retrieved documents. Our first contribution adopts a graph-theoretic perspective to identify a "consistent majority" among retrieved documents to filter out malicious ones. We introduce a novel algorithm based on finding a Maximum Independent Set (MIS) on a document graph where edges encode contradiction. Our MIS variant explicitly prioritizes higher-reliability documents and provides provable robustness guarantees against bounded adversarial corruption under natural assumptions. Recognizing the computational cost of exact MIS for large retrieval sets, our second contribution is a scalable weighted sample and aggregate framework. It explicitly utilizes reliability information, preserving some robustness guarantees while efficiently handling many documents. We present empirical results showing ReliabilityRAG provides superior robustness against adversarial attacks compared to prior methods, maintains high benign accuracy, and excels in long-form generation tasks where prior robustness-focused methods struggled. Our work is a significant step towards more effective, provably robust defenses against retrieved corpus corruption in RAG.

Anonymization of voice seeks to conceal the identity of the speaker while maintaining the utility of speech data. However, residual speaker cues often persist, which pose privacy risks. We propose SegReConcat, a data augmentation method for attacker-side enhancement of automatic speaker verification systems. SegReConcat segments anonymized speech at the word level, rearranges segments using random or similarity-based strategies to disrupt long-term contextual cues, and concatenates them with the original utterance, allowing an attacker to learn source speaker traits from multiple perspectives. The proposed method has been evaluated in the VoicePrivacy Attacker Challenge 2024 framework across seven anonymization systems, SegReConcat improves de-anonymization on five out of seven systems.

Neural speech editing advancements have raised concerns about their misuse in spoofing attacks. Traditional partially edited speech corpora primarily focus on cut-and-paste edits, which, while maintaining speaker consistency, often introduce detectable discontinuities. Recent methods, like A\textsuperscript{3}T and Voicebox, improve transitions by leveraging contextual information. To foster spoofing detection research, we introduce the Speech INfilling Edit (SINE) dataset, created with Voicebox. We detailed the process of re-implementing Voicebox training and dataset creation. Subjective evaluations confirm that speech edited using this novel technique is more challenging to detect than conventional cut-and-paste methods. Despite human difficulty, experimental results demonstrate that self-supervised-based detectors can achieve remarkable performance in detection, localization, and generalization across different edit methods. The dataset and related models will be made publicly available.