The rapid progress of generative adversarial networks (GANs) and diffusion models has enabled the creation of synthetic faces that are increasingly difficult to distinguish from real images. This progress, however, has also amplified the risks of misinformation, fraud, and identity abuse, underscoring the urgent need for detectors that remain robust across diverse generative models. In this work, we introduce Counterfeit Image Pattern High-level Examination via Representation(CIPHER), a deepfake detection framework that systematically reuses and fine-tunes discriminators originally trained for image generation. By extracting scale-adaptive features from ProGAN discriminators and temporal-consistency features from diffusion models, CIPHER captures generation-agnostic artifacts that conventional detectors often overlook. Through extensive experiments across nine state-of-the-art generative models, CIPHER demonstrates superior cross-model detection performance, achieving up to 74.33% F1-score and outperforming existing ViT-based detectors by over 30% in F1-score on average. Notably, our approach maintains robust performance on challenging datasets where baseline methods fail, with up to 88% F1-score on CIFAKE compared to near-zero performance from conventional detectors. These results validate the effectiveness of discriminator reuse and cross-model fine-tuning, establishing CIPHER as a promising approach toward building more generalizable and robust deepfake detection systems in an era of rapidly evolving generative technologies.

Existing jailbreak defence frameworks for Large Vision-Language Models often suffer from a safety utility tradeoff, where strengthening safety inadvertently degrades performance on general visual-grounded reasoning tasks. In this work, we investigate whether safety and utility are inherently antagonistic objectives. We focus on a modality induced bias direction consistently observed across datasets, which arises from suboptimal coupling between the Large Language Model backbone and visual encoders. We further demonstrate that this direction undermines performance on both tasks. Leveraging this insight, we propose Two Birds, One Projection, an efficient inference time jailbreak defence that projects cross-modal features onto the null space of the identified bias direction to remove the corresponding components. Requiring only a single forward pass, our method effectively breaks the conventional tradeoff, simultaneously improving both safety and utility across diverse benchmarks.

AI agents that combine large language models with non-AI system components are rapidly emerging in real-world applications, offering unprecedented automation and flexibility. However, this unprecedented flexibility introduces complex security challenges fundamentally different from those in traditional software systems. This paper presents the first systematic and comprehensive survey of AI agent security, including an analysis of the design space, attack landscape, and defense mechanisms for secure AI agent systems. We further conduct case studies to point out existing gaps in securing agentic AI systems and identify open challenges in this emerging domain. Our work also introduces the first systematic framework for understanding the security risks and defense strategies of AI agents, serving as a foundation for building both secure agentic systems and advancing research in this critical area.

AI agents equipped with tool-calling capabilities are susceptible to Indirect Prompt Injection (IPI) attacks. In this attack scenario, malicious commands hidden within untrusted content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the over-defense dilemma: they deploy expensive, always-on sanitization regardless of actual threat, thereby degrading utility and latency even in benign scenarios. We revisit IPI through a causal ablation perspective: a successful injection manifests as a dominance shift where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment, such as a retrieved document or tool output, provides disproportionate attributable influence. Based on this signature, we propose CausalArmor, a selective defense framework that (i) computes lightweight, leave-one-out ablation-based attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs retroactive Chain-of-Thought masking to prevent the agent from acting on ``poisoned'' reasoning traces. We present a theoretical analysis showing that sanitization based on attribution margins conditionally yields an exponentially small upper bound on the probability of selecting malicious actions. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses while improving explainability and preserving utility and latency of AI agents.

Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly integrated into clinical workflows; however, prompt injection attacks can steer these systems toward clinically unsafe or misleading outputs. We introduce the Medical Prompt Injection Benchmark (MPIB), a dataset-and-benchmark suite for evaluating clinical safety under both direct prompt injection and indirect, RAG-mediated injection across clinically grounded tasks. MPIB emphasizes outcome-level risk via the Clinical Harm Event Rate (CHER), which measures high-severity clinical harm events under a clinically grounded taxonomy, and reports CHER alongside Attack Success Rate (ASR) to disentangle instruction compliance from downstream patient risk. The benchmark comprises 9,697 curated instances constructed through multi-stage quality gates and clinical safety linting. Evaluating MPIB across a diverse set of baseline LLMs and defense configurations, we find that ASR and CHER can diverge substantially, and that robustness depends critically on whether adversarial instructions appear in the user query or in retrieved context. We release MPIB with evaluation code, adversarial baselines, and comprehensive documentation to support reproducible and systematic research on clinical prompt injection. Code and data are available at GitHub (code) and Hugging Face (data).

Text-to-Image (T2I) diffusion models enable high-quality open-ended synthesis, but their real-world deployment demands safeguards that suppress unsafe generations without degrading benign prompt-image alignment. We formalize this tension through a total variation (TV) lens: once the reference conditional distribution is fixed, any nontrivial reduction in unsafe generations necessarily incurs TV deviation from the reference, yielding a principled Safety-Prompt Alignment Trade-off (SPAT). Guided by this view, we propose an inference-only prompt projection framework that selectively intervenes on high-risk prompts via a surrogate objective with verification, mapping them into a tolerance-controlled safe set while leaving benign prompts effectively unchanged, without retraining or fine-tuning the generator. Across four datasets and three diffusion backbones, our approach achieves 16.7-60.0% relative reductions in inappropriate percentage (IP) versus strong model-level alignment baselines, while preserving benign prompt-image alignment on COCO near the unaligned reference.

Modern LLMs are increasingly accessed via black-box APIs, requiring users to transmit sensitive prompts, outputs, and fine-tuning data to external providers, creating a critical privacy risk at the API boundary. We introduce AlienLM, a deployable API-only privacy layer that protects text by translating it into an Alien Language via a vocabulary-scale bijection, enabling lossless recovery on the client side. Using only standard fine-tuning APIs, Alien Adaptation Training (AAT) adapts target models to operate directly on alienized inputs. Across four LLM backbones and seven benchmarks, AlienLM retains over 81\% of plaintext-oracle performance on average, substantially outperforming random-bijection and character-level baselines. Under adversaries with access to model weights, corpus statistics, and learning-based inverse translation, recovery attacks reconstruct fewer than 0.22\% of alienized tokens. Our results demonstrate a practical pathway for privacy-preserving LLM deployment under API-only access, substantially reducing plaintext exposure while maintaining task performance.

The proliferation of hyper-realistic images from Latent Diffusion Models (LDMs) demands robust watermarking, yet existing post-hoc methods are prohibitively slow due to iterative optimization or inversion processes. We introduce PhaseMark, a single-shot, optimization-free framework that directly modulates the phase in the VAE latent frequency domain. This approach makes PhaseMark thousands of times faster than optimization-based techniques while achieving state-of-the-art resilience against severe attacks, including regeneration, without degrading image quality. We analyze four modulation variants, revealing a clear performance-quality trade-off. PhaseMark demonstrates a new paradigm where efficient, resilient watermarking is achieved by exploiting intrinsic latent properties.

Decentralized federated learning (DFL) enables collaborative model training without centralized trust, but it remains vulnerable to Byzantine clients that poison gradients under heterogeneous (Non-IID) data. Existing defenses face a scalability trilemma: distance-based filtering (e.g., Krum) can reject legitimate Non-IID updates, geometric-median methods incur prohibitive $O(n^2 d)$ cost, and many certified defenses are evaluated only on models below 100M parameters. We propose Spectral Sentinel, a Byzantine detection and aggregation framework that leverages a random-matrix-theoretic signature: honest Non-IID gradients produce covariance eigenspectra whose bulk follows the Marchenko-Pastur law, while Byzantine perturbations induce detectable tail anomalies. Our algorithm combines Frequent Directions sketching with data-dependent MP tracking, enabling detection on models up to 1.5B parameters using $O(k^2)$ memory with $k \ll d$. Under a $(σ,f)$ threat model with coordinate-wise honest variance bounded by $σ^2$ and $f < 1/2$ adversaries, we prove $(ε,δ)$-Byzantine resilience with convergence rate $O(σf / \sqrt{T} + f^2 / T)$, and we provide a matching information-theoretic lower bound $Ω(σf / \sqrt{T})$, establishing minimax optimality. We implement the full system with blockchain integration on Polygon networks and validate it across 144 attack-aggregator configurations, achieving 78.4 percent average accuracy versus 48-63 percent for baseline methods.

Recent advancements in diffusion models have made fine-tuning text-to-image models for personalization increasingly accessible, but have also raised significant concerns regarding unauthorized data usage and privacy infringement. Current protection methods are limited to passively degrading image quality, failing to achieve stable control. While Targeted Data Protection (TDP) offers a promising paradigm for active redirection toward user-specified target concepts, existing TDP attempts suffer from poor controllability due to snapshot-matching approaches that fail to account for complete learning dynamics. We introduce TAFAP (Trajectory Alignment via Fine-tuning with Adversarial Perturbations), the first method to successfully achieve effective TDP by controlling the entire training trajectory. Unlike snapshot-based methods whose protective influence is easily diluted as training progresses, TAFAP employs trajectory-matching inspired by dataset distillation to enforce persistent, verifiable transformations throughout fine-tuning. We validate our method through extensive experiments, demonstrating the first successful targeted transformation in diffusion models with simultaneous control over both identity and visual patterns. TAFAP significantly outperforms existing TDP attempts, achieving robust redirection toward target concepts while maintaining high image quality. This work enables verifiable safeguards and provides a new framework for controlling and tracing alterations in diffusion model outputs.

Adversarial examples in neural networks have been extensively studied in Euclidean geometry, but recent advances in \textit{hyperbolic networks} call for a reevaluation of attack strategies in non-Euclidean geometries. Existing methods such as FGSM and PGD apply perturbations without regard to the underlying hyperbolic structure, potentially leading to inefficient or geometrically inconsistent attacks. In this work, we propose a novel adversarial attack that explicitly leverages the geometric properties of hyperbolic space. Specifically, we compute the gradient of the loss function in the tangent space of hyperbolic space, decompose it into a radial (depth) component and an angular (semantic) component, and apply perturbation derived solely from the angular direction. Our method generates adversarial examples by focusing perturbations in semantically sensitive directions encoded in angular movement within the hyperbolic geometry. Empirical results on image classification, cross-modal retrieval tasks and network architectures demonstrate that our attack achieves higher fooling rates than conventional adversarial attacks, while producing high-impact perturbations with deeper insights into vulnerabilities of hyperbolic embeddings. This work highlights the importance of geometry-aware adversarial strategies in curved representation spaces and provides a principled framework for attacking hierarchical embeddings.

While diffusion-based T2I models have achieved remarkable image generation quality, they also enable easy creation of harmful content, raising social concerns and highlighting the need for safer generation. Existing inference-time guiding methods lack both adaptivity--adjusting guidance strength based on the prompt--and selectivity--targeting only unsafe regions of the image. Our method, SP-Guard, addresses these limitations by estimating prompt harmfulness and applying a selective guidance mask to guide only unsafe areas. Experiments show that SP-Guard generates safer images than existing methods while minimizing unintended content alteration. Beyond improving safety, our findings highlight the importance of transparency and controllability in image generation.

One of the key challenges of modern AI models is ensuring that they provide helpful responses to benign queries while refusing malicious ones. But often, the models are vulnerable to multimodal queries with harmful intent embedded in images. One approach for safety alignment is training with extensive safety datasets at the significant costs in both dataset curation and training. Inference-time alignment mitigates these costs, but introduces two drawbacks: excessive refusals from misclassified benign queries and slower inference speed due to iterative output adjustments. To overcome these limitations, we propose to reformulate queries to strengthen cross-modal attention to safety-critical image regions, enabling accurate risk assessment at the query level. Using the assessed risk, it adaptively steers activations to generate responses that are safe and helpful without overhead from iterative output adjustments. We call this Risk-adaptive Activation Steering (RAS). Extensive experiments across multiple benchmarks on multimodal safety and utility demonstrate that the RAS significantly reduces attack success rates, preserves general task performance, and improves inference speed over prior inference-time defenses.

The promise of LLM watermarking rests on a core assumption that a specific watermark proves authorship by a specific model. We demonstrate that this assumption is dangerously flawed. We introduce the threat of watermark spoofing, a sophisticated attack that allows a malicious model to generate text containing the authentic-looking watermark of a trusted, victim model. This enables the seamless misattribution of harmful content, such as disinformation, to reputable sources. The key to our attack is repurposing watermark radioactivity, the unintended inheritance of data patterns during fine-tuning, from a discoverable trait into an attack vector. By distilling knowledge from a watermarked teacher model, our framework allows an attacker to steal and replicate the watermarking signal of the victim model. This work reveals a critical security gap in text authorship verification and calls for a paradigm shift towards technologies capable of distinguishing authentic watermarks from expertly imitated ones. Our code is available at https://github.com/hsannn/ditto.git.

Despite their impressive generative capabilities, text-to-image diffusion models often memorize and replicate training data, prompting serious concerns over privacy and copyright. Recent work has attributed this memorization to an attraction basin-a region where applying classifier-free guidance (CFG) steers the denoising trajectory toward memorized outputs-and has proposed deferring CFG application until the denoising trajectory escapes this basin. However, such delays often result in non-memorized images that are poorly aligned with the input prompts, highlighting the need to promote earlier escape so that CFG can be applied sooner in the denoising process. In this work, we show that the initial noise sample plays a crucial role in determining when this escape occurs. We empirically observe that different initial samples lead to varying escape times. Building on this insight, we propose two mitigation strategies that adjust the initial noise-either collectively or individually-to find and utilize initial samples that encourage earlier basin escape. These approaches significantly reduce memorization while preserving image-text alignment.

Along with the rapid advancement of numerous Text-to-Video (T2V) models, growing concerns have emerged regarding their safety risks. While recent studies have explored vulnerabilities in models like LLMs, VLMs, and Text-to-Image (T2I) models through jailbreak attacks, T2V models remain largely unexplored, leaving a significant safety gap. To address this gap, we introduce SceneSplit, a novel black-box jailbreak method that works by fragmenting a harmful narrative into multiple scenes, each individually benign. This approach manipulates the generative output space, the abstract set of all potential video outputs for a given prompt, using the combination of scenes as a powerful constraint to guide the final outcome. While each scene individually corresponds to a wide and safe space where most outcomes are benign, their sequential combination collectively restricts this space, narrowing it to an unsafe region and significantly increasing the likelihood of generating a harmful video. This core mechanism is further enhanced through iterative scene manipulation, which bypasses the safety filter within this constrained unsafe region. Additionally, a strategy library that reuses successful attack patterns further improves the attack's overall effectiveness and robustness. To validate our method, we evaluate SceneSplit across 11 safety categories on T2V models. Our results show that it achieves a high average Attack Success Rate (ASR) of 77.2% on Luma Ray2, 84.1% on Hailuo, and 78.2% on Veo2, significantly outperforming the existing baseline. Through this work, we demonstrate that current T2V safety mechanisms are vulnerable to attacks that exploit narrative structure, providing new insights for understanding and improving the safety of T2V models.

Large language models trained on web-scale data can memorize private or sensitive knowledge, raising significant privacy risks. Although some unlearning methods mitigate these risks, they remain vulnerable to "relearning" during subsequent training, allowing a substantial portion of forgotten knowledge to resurface. In this paper, we show that widely used unlearning methods cause shallow alignment: instead of faithfully erasing target knowledge, they generate spurious unlearning neurons that amplify negative influence to hide it. To overcome this limitation, we introduce Ssiuu, a new class of unlearning methods that employs attribution-guided regularization to prevent spurious negative influence and faithfully remove target knowledge. Experimental results confirm that our method reliably erases target knowledge and outperforms strong baselines across two practical retraining scenarios: (1) adversarial injection of private data, and (2) benign attack using an instruction-following benchmark. Our findings highlight the necessity of robust and faithful unlearning methods for safe deployment of language models.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, but their potential misuse for harmful purposes remains a significant concern. To strengthen defenses against such vulnerabilities, it is essential to investigate universal jailbreak attacks that exploit intrinsic weaknesses in the architecture and learning paradigms of LLMs. In response, we propose \textbf{H}armful \textbf{P}rompt \textbf{La}undering (HaPLa), a novel and broadly applicable jailbreaking technique that requires only black-box access to target models. HaPLa incorporates two primary strategies: 1) \textit{abductive framing}, which instructs LLMs to infer plausible intermediate steps toward harmful activities, rather than directly responding to explicit harmful queries; and 2) \textit{symbolic encoding}, a lightweight and flexible approach designed to obfuscate harmful content, given that current LLMs remain sensitive primarily to explicit harmful keywords. Experimental results show that HaPLa achieves over 95% attack success rate on GPT-series models and 70% across all targets. Further analysis with diverse symbolic encoding rules also reveals a fundamental challenge: it remains difficult to safely tune LLMs without significantly diminishing their helpfulness in responding to benign queries.

In-context learning (ICL) in Large Language Models (LLMs) has shown remarkable performance across various tasks without requiring fine-tuning. However, recent studies have highlighted the risk of private data leakage through the prompt in ICL, especially when LLMs are exposed to malicious attacks. While differential privacy (DP) provides strong privacy guarantees, it often significantly reduces the utility of in-context learning (ICL). To address this challenge, we incorporate task-related public data into the ICL framework while maintaining the DP guarantee. Based on this approach, we propose a private in-context learning algorithm that effectively balances privacy protection and model utility. Through experiments, we demonstrate that our approach significantly improves the utility of private ICL with the assistance of public data. Additionally, we show that our method is robust against membership inference attacks, demonstrating empirical privacy protection.

Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at https://github.com/thomas11809/SFWMark