Existing research on LLM agent security mainly focuses on prompt injection and unsafe input/output behaviors. However, as agents increasingly rely on third-party tools and MCP servers, a new class of supply-chain threats has emerged, where malicious behaviors are embedded in seemingly benign tools, silently hijacking agent execution, leaking sensitive data, or triggering unauthorized actions. Despite their growing impact, there is currently no comprehensive benchmark for evaluating such threats. To bridge this gap, we introduce SC-Inject-Bench, a large-scale benchmark comprising over 10,000 malicious MCP tools grounded in a taxonomy of 25+ attack types derived from MITRE ATT&CK targeting supply-chain threats. We observe that existing MCP scanners and semantic guardrails perform poorly on this benchmark. Motivated by this finding, we propose ShieldNet, a network-level guardrail framework that detects supply-chain poisoning by observing real network interactions rather than surface-level tool traces. ShieldNet integrates a man-in-the-middle (MITM) proxy and an event extractor to identify critical network behaviors, which are then processed by a lightweight classifier for attack detection. Extensive experiments show that ShieldNet achieves strong detection performance (up to 0.995 F-1 with only 0.8% false positives) while introducing little runtime overhead, substantially outperforming existing MCP scanners and LLM-based guardrails.

Large language models (LLMs) are increasingly deployed for everyday tasks, including food preparation and health-related guidance. However, food safety remains a high-stakes domain where inaccurate or misleading information can cause severe real-world harm. Despite these risks, current LLMs and safety guardrails lack rigorous alignment tailored to domain-specific food hazards. To address this gap, we introduce FoodGuardBench, the first comprehensive benchmark comprising 3,339 queries grounded in FDA guidelines, designed to evaluate the safety and robustness of LLMs. By constructing a taxonomy of food safety principles and employing representative jailbreak attacks (e.g., AutoDAN and PAP), we systematically evaluate existing LLMs and guardrails. Our evaluation results reveal three critical vulnerabilities: First, current LLMs exhibit sparse safety alignment in the food-related domain, easily succumbing to a few canonical jailbreak strategies. Second, when compromised, LLMs frequently generate actionable yet harmful instructions, inadvertently empowering malicious actors and posing tangible risks. Third, existing LLM-based guardrails systematically overlook these domain-specific threats, failing to detect a substantial volume of malicious inputs. To mitigate these vulnerabilities, we introduce FoodGuard-4B, a specialized guardrail model fine-tuned on our datasets to safeguard LLMs within food-related domains.

Safety alignment aims to ensure that large language models (LLMs) refuse harmful requests by post-training on harmful queries paired with refusal answers. Although safety alignment is widely adopted in industry, the overrefusal problem where aligned LLMs also reject benign queries after safety alignment post-training, remains insufficiently studied. Such an issue degrades the usability of safety alignment in real-world applications. In this paper, we examine how overrefusal arises under safety alignment, and propose a mitigation strategy inspired by our findings. We define refusal triggers as linguistic cues in the training data that elicit refusal responses, safety alignment encourages LLMs to associate refusal triggers within a training sample with refusal responses, leading aligned LLMs to refuse harmful queries. However, the refusal triggers include not only harmful linguistic cues but also non-harmful cues, therefore causing overrefusal to benign queries. Building on this mechanistic analysis, we propose a method that explicitly considers refusal triggers in the safety alignment fine-tuning. Empirical results demonstrate that our approach achieves a more favorable trade-off between defense against jailbreak attacks and responsiveness to benign queries, outperforming prior methods. Warning: this paper contains harmful and biased sentences.

Indirect prompt injection threatens LLM agents by embedding malicious instructions in external content, enabling unauthorized actions and data theft. LLM agents maintain working memory through their context window, which stores interaction history for decision-making. Conventional agents indiscriminately accumulate all tool outputs and reasoning traces in this memory, creating two critical vulnerabilities: (1) injected instructions persist throughout the workflow, granting attackers multiple opportunities to manipulate behavior, and (2) verbose, non-essential content degrades decision-making capabilities. Existing defenses treat bloated memory as given and focus on remaining resilient, rather than reducing unnecessary accumulation to prevent the attack. We present AgentSys, a framework that defends against indirect prompt injection through explicit memory management. Inspired by process memory isolation in operating systems, AgentSys organizes agents hierarchically: a main agent spawns worker agents for tool calls, each running in an isolated context and able to spawn nested workers for subtasks. External data and subtask traces never enter the main agent's memory; only schema-validated return values can cross boundaries through deterministic JSON parsing. Ablations show isolation alone cuts attack success to 2.19%, and adding a validator/sanitizer further improves defense with event-triggered checks whose overhead scales with operations rather than context length. On AgentDojo and ASB, AgentSys achieves 0.78% and 4.25% attack success while slightly improving benign utility over undefended baselines. It remains robust to adaptive attackers and across multiple foundation models, showing that explicit memory management enables secure, dynamic LLM agent architectures. Our code is available at: https://github.com/ruoyaow/agentsys-memory.

AI agents that autonomously interact with external tools and environments show great promise across real-world applications. However, the external data which agent consumes also leads to the risk of indirect prompt injection attacks, where malicious instructions embedded in third-party content hijack agent behavior. Guided by benchmarks, such as AgentDojo, there has been significant amount of progress in developing defense against the said attacks. As the technology continues to mature, and that agents are increasingly being relied upon for more complex tasks, there is increasing pressing need to also evolve the benchmark to reflect threat landscape faced by emerging agentic systems. In this work, we reveal three fundamental flaws in current benchmarks and push the frontier along these dimensions: (i) lack of dynamic open-ended tasks, (ii) lack of helpful instructions, and (iii) simplistic user tasks. To bridge this gap, we introduce AgentDyn, a manually designed benchmark featuring 60 challenging open-ended tasks and 560 injection test cases across Shopping, GitHub, and Daily Life. Unlike prior static benchmarks, AgentDyn requires dynamic planning and incorporates helpful third-party instructions. Our evaluation of ten state-of-the-art defenses suggests that almost all existing defenses are either not secure enough or suffer from significant over-defense, revealing that existing defenses are still far from real-world deployment. Our benchmark is available at https://github.com/leolee99/AgentDyn.

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.

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.

Large language models (LLMs) are increasingly deployed in security-sensitive applications, where they must follow system- or developer-specified instructions that define the intended task behavior, while completing benign user requests. When adversarial instructions appear in user queries or externally retrieved content, models may override intended logic. Recent defenses rely on supervised fine-tuning with benign and malicious labels. Although these methods achieve high attack rejection rates, we find that they rely on narrow correlations in defense data rather than harmful intent, leading to systematic rejection of safe inputs. We analyze three recurring shortcut behaviors induced by defense fine-tuning. \emph{Position bias} arises when benign content placed later in a prompt is rejected at much higher rates; across reasoning benchmarks, suffix-task rejection rises from below \textbf{10\%} to as high as \textbf{90\%}. \emph{Token trigger bias} occurs when strings common in attack data raise rejection probability even in benign contexts; inserting a single trigger token increases false refusals by up to \textbf{50\%}. \emph{Topic generalization bias} reflects poor generalization beyond the defense data distribution, with defended models suffering test-time accuracy drops of up to \textbf{40\%}. These findings suggest that current prompt-injection defenses frequently respond to attack-like surface patterns rather than the underlying intent. We introduce controlled diagnostic datasets and a systematic evaluation across two base models and multiple defense pipelines, highlighting limitations of supervised fine-tuning for reliable LLM security.

Deepfake detection models have achieved high accuracy in identifying synthetic media, but their decision processes remain largely opaque. In this paper we present a mechanistic interpretability framework for deepfake detection applied to a vision-language model. Our approach combines a sparse autoencoder (SAE) analysis of internal network representations with a novel forensic manifold analysis that probes how the model's features respond to controlled forensic artifact manipulations. We demonstrate that only a small fraction of latent features are actively used in each layer, and that the geometric properties of the model's feature manifold, including intrinsic dimensionality, curvature, and feature selectivity, vary systematically with different types of deepfake artifacts. These insights provide a first step toward opening the "black box" of deepfake detectors, allowing us to identify which learned features correspond to specific forensic artifacts and to guide the development of more interpretable and robust models.

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).

Recent advancements in large-scale generative models have enabled the creation of high-quality images and videos, but have also raised significant safety concerns regarding the generation of unsafe content. To mitigate this, concept erasure methods have been developed to remove undesirable concepts from pre-trained models. However, existing methods remain vulnerable to adversarial attacks that can regenerate the erased content. Moreover, achieving robust erasure often degrades the model's generative quality for safe, unrelated concepts, creating a difficult trade-off between safety and performance. To address this challenge, we introduce Classifier-Guided Concept Erasure (CGCE), an efficient plug-and-play framework that provides robust concept erasure for diverse generative models without altering their original weights. CGCE uses a lightweight classifier operating on text embeddings to first detect and then refine prompts containing undesired concepts. This approach is highly scalable, allowing for multi-concept erasure by aggregating guidance from several classifiers. By modifying only unsafe embeddings at inference time, our method prevents harmful content generation while preserving the model's original quality on benign prompts. Extensive experiments show that CGCE achieves state-of-the-art robustness against a wide range of red-teaming attacks. Our approach also maintains high generative utility, demonstrating a superior balance between safety and performance. We showcase the versatility of CGCE through its successful application to various modern T2I and T2V models, establishing it as a practical and effective solution for safe generative AI.

Computer-use agent (CUA) frameworks, powered by large language models (LLMs) or multimodal LLMs (MLLMs), are rapidly maturing as assistants that can perceive context, reason, and act directly within software environments. Among their most critical applications is operating system (OS) control. As CUAs in the OS domain become increasingly embedded in daily operations, it is imperative to examine their real-world security implications, specifically whether CUAs can be misused to perform realistic, security-relevant attacks. Existing works exhibit four major limitations: Missing attacker-knowledge model on tactics, techniques, and procedures (TTP), Incomplete coverage for end-to-end kill chains, unrealistic environment without multi-host and encrypted user credentials, and unreliable judgment dependent on LLM-as-a-Judge. To address these gaps, we propose AdvCUA, the first benchmark aligned with real-world TTPs in MITRE ATT&CK Enterprise Matrix, which comprises 140 tasks, including 40 direct malicious tasks, 74 TTP-based malicious tasks, and 26 end-to-end kill chains, systematically evaluates CUAs under a realistic enterprise OS security threat in a multi-host environment sandbox by hard-coded evaluation. We evaluate the existing five mainstream CUAs, including ReAct, AutoGPT, Gemini CLI, Cursor CLI, and Cursor IDE based on 8 foundation LLMs. The results demonstrate that current frontier CUAs do not adequately cover OS security-centric threats. These capabilities of CUAs reduce dependence on custom malware and deep domain expertise, enabling even inexperienced attackers to mount complex enterprise intrusions, which raises social concern about the responsibility and security of CUAs.

Recent advancements in jailbreaking large language models (LLMs), such as AutoDAN-Turbo, have demonstrated the power of automated strategy discovery. AutoDAN-Turbo employs a lifelong learning agent to build a rich library of attack strategies from scratch. While highly effective, its test-time generation process involves sampling a strategy and generating a single corresponding attack prompt, which may not fully exploit the potential of the learned strategy library. In this paper, we propose to further improve the attack performance of AutoDAN-Turbo through test-time scaling. We introduce two distinct scaling methods: Best-of-N and Beam Search. The Best-of-N method generates N candidate attack prompts from a sampled strategy and selects the most effective one based on a scorer model. The Beam Search method conducts a more exhaustive search by exploring combinations of strategies from the library to discover more potent and synergistic attack vectors. According to the experiments, the proposed methods significantly boost performance, with Beam Search increasing the attack success rate by up to 15.6 percentage points on Llama-3.1-70B-Instruct and achieving a nearly 60% relative improvement against the highly robust GPT-o4-mini compared to the vanilla method.

Large Language Models (LLMs) and their multimodal extensions are becoming increasingly popular. One common approach to enable multimodality is to cascade domain-specific encoders with an LLM, making the resulting model inherit vulnerabilities from all of its components. In this work, we present the first systematic study of audio backdoor attacks against speech language models. We demonstrate its effectiveness across four speech encoders and three datasets, covering four tasks: automatic speech recognition (ASR), speech emotion recognition, and gender and age prediction. The attack consistently achieves high success rates, ranging from 90.76% to 99.41%. To better understand how backdoors propagate, we conduct a component-wise analysis to identify the most vulnerable stages of the pipeline. Finally, we propose a fine-tuning-based defense that mitigates the threat of poisoned pretrained encoders.

Embodied Artificial Intelligence (AI) promises to handle edge cases in robotic vehicle systems where data is scarce by using common-sense reasoning grounded in perception and action to generalize beyond training distributions and adapt to novel real-world situations. These capabilities, however, also create new security risks. In this paper, we introduce CHAI (Command Hijacking against embodied AI), a physical environment indirect prompt injection attack that exploits the multimodal language interpretation abilities of AI models. CHAI embeds deceptive natural language instructions, such as misleading signs, in visual input, systematically searches the token space, builds a dictionary of prompts, and guides an attacker model to generate Visual Attack Prompts. We evaluate CHAI on four LVLM agents: drone emergency landing, autonomous driving, aerial object tracking, and on a real robotic vehicle. Our experiments show that CHAI consistently outperforms state-of-the-art attacks. By exploiting the semantic and multimodal reasoning strengths of next-generation embodied AI systems, CHAI underscores the urgent need for defenses that extend beyond traditional adversarial robustness.

Current anti-spoofing systems remain vulnerable to expressive and emotional synthetic speech, since they rarely leverage prosody as a discriminative cue. Prosody is central to human expressiveness and emotion, and humans instinctively use prosodic cues such as F0 patterns and voiced/unvoiced structure to distinguish natural from synthetic speech. In this paper, we propose HuLA, a two-stage prosody-aware multi-task learning framework for spoof detection. In Stage 1, a self-supervised learning (SSL) backbone is trained on real speech with auxiliary tasks of F0 prediction and voiced/unvoiced classification, enhancing its ability to capture natural prosodic variation similar to human perceptual learning. In Stage 2, the model is jointly optimized for spoof detection and prosody tasks on both real and synthetic data, leveraging prosodic awareness to detect mismatches between natural and expressive synthetic speech. Experiments show that HuLA consistently outperforms strong baselines on challenging out-of-domain dataset, including expressive, emotional, and cross-lingual attacks. These results demonstrate that explicit prosodic supervision, combined with SSL embeddings, substantially improves robustness against advanced synthetic speech attacks.

Multimodal pre-trained models (e.g., ImageBind), which align distinct data modalities into a shared embedding space, have shown remarkable success across downstream tasks. However, their increasing adoption raises serious security concerns, especially regarding targeted adversarial attacks. In this paper, we show that existing targeted adversarial attacks on multimodal pre-trained models still have limitations in two aspects: generalizability and undetectability. Specifically, the crafted targeted adversarial examples (AEs) exhibit limited generalization to partially known or semantically similar targets in cross-modal alignment tasks (i.e., limited generalizability) and can be easily detected by simple anomaly detection methods (i.e., limited undetectability). To address these limitations, we propose a novel method called Proxy Targeted Attack (PTA), which leverages multiple source-modal and target-modal proxies to optimize targeted AEs, ensuring they remain evasive to defenses while aligning with multiple potential targets. We also provide theoretical analyses to highlight the relationship between generalizability and undetectability and to ensure optimal generalizability while meeting the specified requirements for undetectability. Furthermore, experimental results demonstrate that our PTA can achieve a high success rate across various related targets and remain undetectable against multiple anomaly detection methods.

Different from traditional sentence-level audio deepfake detection (ADD), partial audio deepfake detection (PADD) requires frame-level positioning of the location of fake speech. While some progress has been made in this area, leveraging semantic information from audio, especially named entities, remains an underexplored aspect. To this end, we propose NE-PADD, a novel method for Partial Audio Deepfake Detection (PADD) that leverages named entity knowledge through two parallel branches: Speech Name Entity Recognition (SpeechNER) and PADD. The approach incorporates two attention aggregation mechanisms: Attention Fusion (AF) for combining attention weights and Attention Transfer (AT) for guiding PADD with named entity semantics using an auxiliary loss. Built on the PartialSpoof-NER dataset, experiments show our method outperforms existing baselines, proving the effectiveness of integrating named entity knowledge in PADD. The code is available at https://github.com/AI-S2-Lab/NE-PADD.

In this work, we propose a multi-target backdoor attack against speaker identification using position-independent clicking sounds as triggers. Unlike previous single-target approaches, our method targets up to 50 speakers simultaneously, achieving success rates of up to 95.04%. To simulate more realistic attack conditions, we vary the signal-to-noise ratio between speech and trigger, demonstrating a trade-off between stealth and effectiveness. We further extend the attack to the speaker verification task by selecting the most similar training speaker - based on cosine similarity - as a proxy target. The attack is most effective when target and enrolled speaker pairs are highly similar, reaching success rates of up to 90% in such cases.

Audio-visual temporal deepfake localization under the content-driven partial manipulation remains a highly challenging task. In this scenario, the deepfake regions are usually only spanning a few frames, with the majority of the rest remaining identical to the original. To tackle this, we propose a Hierarchical Boundary Modeling Network (HBMNet), which includes three modules: an Audio-Visual Feature Encoder that extracts discriminative frame-level representations, a Coarse Proposal Generator that predicts candidate boundary regions, and a Fine-grained Probabilities Generator that refines these proposals using bidirectional boundary-content probabilities. From the modality perspective, we enhance audio-visual learning through dedicated encoding and fusion, reinforced by frame-level supervision to boost discriminability. From the temporal perspective, HBMNet integrates multi-scale cues and bidirectional boundary-content relationships. Experiments show that encoding and fusion primarily improve precision, while frame-level supervision boosts recall. Each module (audio-visual fusion, temporal scales, bi-directionality) contributes complementary benefits, collectively enhancing localization performance. HBMNet outperforms BA-TFD and UMMAFormer and shows improved potential scalability with more training data.