Backdoor mechanisms have traditionally been studied as security threats that compromise the integrity of machine learning models. However, the same mechanism -- the conditional activation of specific behaviors through input triggers -- can also serve as a controllable and auditable interface for trustworthy model behavior. In this work, we present \textbf{Backdoor4Good (B4G)}, a unified benchmark and framework for \textit{beneficial backdoor} applications in large language models (LLMs). Unlike conventional backdoor studies focused on attacks and defenses, B4G repurposes backdoor conditioning for Beneficial Tasks that enhance safety, controllability, and accountability. It formalizes beneficial backdoor learning under a triplet formulation $(T, A, U)$, representing the \emph{Trigger}, \emph{Activation mechanism}, and \emph{Utility function}, and implements a benchmark covering four trust-centric applications. Through extensive experiments across Llama3.1-8B, Gemma-2-9B, Qwen2.5-7B, and Llama2-13B, we show that beneficial backdoors can achieve high controllability, tamper-resistance, and stealthiness while preserving clean-task performance. Our findings demonstrate new insights that backdoors need not be inherently malicious; when properly designed, they can serve as modular, interpretable, and beneficial building blocks for trustworthy AI systems. Our code and datasets are available at https://github.com/bboylyg/BackdoorLLM/B4G.

Multi-turn jailbreak attacks are effective against text-only large language models (LLMs) by gradually introducing malicious content across turns. When extended to large vision-language models (LVLMs), we find that naively adding visual inputs can cause existing multi-turn jailbreaks to be easily defended. For example, overly malicious visual input will easily trigger the defense mechanism of safety-aligned LVLMs, making the response more conservative. To address this, we propose MAPA: a multi-turn adaptive prompting attack that 1) at each turn, alternates text-vision attack actions to elicit the most malicious response; and 2) across turns, adjusts the attack trajectory through iterative back-and-forth refinement to gradually amplify response maliciousness. This two-level design enables MAPA to consistently outperform state-of-the-art methods, improving attack success rates by 11-35% on recent benchmarks against LLaVA-V1.6-Mistral-7B, Qwen2.5-VL-7B-Instruct, Llama-3.2-Vision-11B-Instruct and GPT-4o-mini.

Recent studies have shown that CLIP model's adversarial robustness in zero-shot classification tasks can be enhanced by adversarially fine-tuning its image encoder with adversarial examples (AEs), which are generated by minimizing the cosine similarity between images and a hand-crafted template (e.g., ''A photo of a {label}''). However, it has been shown that the cosine similarity between a single image and a single hand-crafted template is insufficient to measure the similarity for image-text pairs. Building on this, in this paper, we find that the AEs generated using cosine similarity may fail to fool CLIP when the similarity metric is replaced with semantically enriched alternatives, making the image encoder fine-tuned with these AEs less robust. To overcome this issue, we first propose a semantic-ensemble attack to generate semantic-aware AEs by minimizing the average similarity between the original image and an ensemble of refined textual descriptions. These descriptions are initially generated by a foundation model to capture core semantic features beyond hand-crafted templates and are then refined to reduce hallucinations. To this end, we propose Semantic-aware Adversarial Fine-Tuning (SAFT), which fine-tunes CLIP's image encoder with semantic-aware AEs. Extensive experiments show that SAFT outperforms current methods, achieving substantial improvements in zero-shot adversarial robustness across 16 datasets. Our code is available at: https://github.com/tmlr-group/SAFT.

Code models are increasingly adopted in software development but remain vulnerable to backdoor attacks via poisoned training data. Existing backdoor attacks on code models face a fundamental trade-off between transferability and stealthiness. Static trigger-based attacks insert fixed dead code patterns that transfer well across models and datasets but are easily detected by code-specific defenses. In contrast, dynamic trigger-based attacks adaptively generate context-aware triggers to evade detection but suffer from poor cross-dataset transferability. Moreover, they rely on unrealistic assumptions of identical data distributions between poisoned and victim training data, limiting their practicality. To overcome these limitations, we propose Sharpness-aware Transferable Adversarial Backdoor (STAB), a novel attack that achieves both transferability and stealthiness without requiring complete victim data. STAB is motivated by the observation that adversarial perturbations in flat regions of the loss landscape transfer more effectively across datasets than those in sharp minima. To this end, we train a surrogate model using Sharpness-Aware Minimization to guide model parameters toward flat loss regions, and employ Gumbel-Softmax optimization to enable differentiable search over discrete trigger tokens for generating context-aware adversarial triggers. Experiments across three datasets and two code models show that STAB outperforms prior attacks in terms of transferability and stealthiness. It achieves a 73.2% average attack success rate after defense, outperforming static trigger-based attacks that fail under defense. STAB also surpasses the best dynamic trigger-based attack by 12.4% in cross-dataset attack success rate and maintains performance on clean inputs.

Vision-language models (VLMs) extend large language models (LLMs) with vision encoders, enabling text generation conditioned on both images and text. However, this multimodal integration expands the attack surface by exposing the model to image-based jailbreaks crafted to induce harmful responses. Existing gradient-based jailbreak methods transfer poorly, as adversarial patterns overfit to a single white-box surrogate and fail to generalise to black-box models. In this work, we propose Universal and transferable jailbreak (UltraBreak), a framework that constrains adversarial patterns through transformations and regularisation in the vision space, while relaxing textual targets through semantic-based objectives. By defining its loss in the textual embedding space of the target LLM, UltraBreak discovers universal adversarial patterns that generalise across diverse jailbreak objectives. This combination of vision-level regularisation and semantically guided textual supervision mitigates surrogate overfitting and enables strong transferability across both models and attack targets. Extensive experiments show that UltraBreak consistently outperforms prior jailbreak methods. Further analysis reveals why earlier approaches fail to transfer, highlighting that smoothing the loss landscape via semantic objectives is crucial for enabling universal and transferable jailbreaks. The code is publicly available in our \href{https://github.com/kaiyuanCui/UltraBreak}{GitHub repository}.

Autonomous code agents built on large language models are reshaping software and AI development through tool use, long-horizon reasoning, and self-directed interaction. However, this autonomy introduces a previously unrecognized security risk: agentic interaction fundamentally expands the LLM attack surface, enabling systematic probing and recovery of hidden system prompts that guide model behavior. We identify system prompt extraction as an emergent vulnerability intrinsic to code agents and present \textbf{\textsc{JustAsk}}, a self-evolving framework that autonomously discovers effective extraction strategies through interaction alone. Unlike prior prompt-engineering or dataset-based attacks, \textsc{JustAsk} requires no handcrafted prompts, labeled supervision, or privileged access beyond standard user interaction. It formulates extraction as an online exploration problem, using Upper Confidence Bound-based strategy selection and a hierarchical skill space spanning atomic probes and high-level orchestration. These skills exploit imperfect system-instruction generalization and inherent tensions between helpfulness and safety. Evaluated on \textbf{41} black-box commercial models across multiple providers, \textsc{JustAsk} consistently achieves full or near-complete system prompt recovery, revealing recurring design- and architecture-level vulnerabilities. Our results expose system prompts as a critical yet largely unprotected attack surface in modern agent systems.

Protecting the intellectual property of large language models requires robust ownership verification. Conventional backdoor fingerprinting, however, is flawed by a stealth-robustness paradox: to be robust, these methods force models to memorize fixed responses to high-perplexity triggers, but this targeted overfitting creates detectable statistical artifacts. We resolve this paradox with KinGuard, a framework that embeds a private knowledge corpus built on structured kinship narratives. Instead of memorizing superficial triggers, the model internalizes this knowledge via incremental pre-training, and ownership is verified by probing its conceptual understanding. Extensive experiments demonstrate KinGuard's superior effectiveness, stealth, and resilience against a battery of attacks including fine-tuning, input perturbation, and model merging. Our work establishes knowledge-based embedding as a practical and secure paradigm for model fingerprinting.

Humans are susceptible to undesirable behaviours and privacy leaks under the influence of alcohol. This paper investigates drunk language, i.e., text written under the influence of alcohol, as a driver for safety failures in large language models (LLMs). We investigate three mechanisms for inducing drunk language in LLMs: persona-based prompting, causal fine-tuning, and reinforcement-based post-training. When evaluated on 5 LLMs, we observe a higher susceptibility to jailbreaking on JailbreakBench (even in the presence of defences) and privacy leaks on ConfAIde, where both benchmarks are in English, as compared to the base LLMs as well as previously reported approaches. Via a robust combination of manual evaluation and LLM-based evaluators and analysis of error categories, our findings highlight a correspondence between human-intoxicated behaviour, and anthropomorphism in LLMs induced with drunk language. The simplicity and efficiency of our drunk language inducement approaches position them as potential counters for LLM safety tuning, highlighting significant risks to LLM safety.

Large Language Models (LLMs) excel at text comprehension and generation, making them ideal for automated tasks like code review and content moderation. However, our research identifies a vulnerability: LLMs can be manipulated by "adversarial instructions" hidden in input data, such as resumes or code, causing them to deviate from their intended task. Notably, while defenses may exist for mature domains such as code review, they are often absent in other common applications such as resume screening and peer review. This paper introduces a benchmark to assess this vulnerability in resume screening, revealing attack success rates exceeding 80% for certain attack types. We evaluate two defense mechanisms: prompt-based defenses achieve 10.1% attack reduction with 12.5% false rejection increase, while our proposed FIDS (Foreign Instruction Detection through Separation) using LoRA adaptation achieves 15.4% attack reduction with 10.4% false rejection increase. The combined approach provides 26.3% attack reduction, demonstrating that training-time defenses outperform inference-time mitigations in both security and utility preservation.

Accurate time-series forecasting is increasingly critical for planning and operations in low-carbon power systems. Emerging time-series large language models (TS-LLMs) now deliver this capability at scale, requiring no task-specific retraining, and are quickly becoming essential components within the Internet-of-Energy (IoE) ecosystem. However, their real-world deployment is complicated by a critical vulnerability: adversarial examples (AEs). Detecting these AEs is challenging because (i) adversarial perturbations are optimized across the entire input sequence and exploit global temporal dependencies, which renders local detection methods ineffective, and (ii) unlike traditional forecasting models with fixed input dimensions, TS-LLMs accept sequences of variable length, increasing variability that complicates detection. To address these challenges, we propose a plug-in detection framework that capitalizes on the TS-LLM's own variable-length input capability. Our method uses sampling-induced divergence as a detection signal. Given an input sequence, we generate multiple shortened variants and detect AEs by measuring the consistency of their forecasts: Benign sequences tend to produce stable predictions under sampling, whereas adversarial sequences show low forecast similarity, because perturbations optimized for a full-length sequence do not transfer reliably to shorter, differently-structured subsamples. We evaluate our approach on three representative TS-LLMs (TimeGPT, TimesFM, and TimeLLM) across three energy datasets: ETTh2 (Electricity Transformer Temperature), NI (Hourly Energy Consumption), and Consumption (Hourly Electricity Consumption and Production). Empirical results confirm strong and robust detection performance across both black-box and white-box attack scenarios, highlighting its practicality as a reliable safeguard for TS-LLM forecasting in real-world energy systems.

3D Gaussian Splatting (3DGS) has enabled the creation of digital assets and downstream applications, underscoring the need for robust copyright protection via digital watermarking. However, existing 3DGS watermarking methods remain highly vulnerable to diffusion-based editing, which can easily erase embedded provenance. This challenge highlights the urgent need for 3DGS watermarking techniques that are intrinsically resilient to diffusion-based editing. In this paper, we introduce RDSplat, a Robust watermarking paradigm against Diffusion editing for 3D Gaussian Splatting. RDSplat embeds watermarks into 3DGS components that diffusion-based editing inherently preserve, achieved through (i) proactively targeting low-frequency Gaussians and (ii) adversarial training with a diffusion proxy. Specifically, we introduce a multi-domain framework that operates natively in 3DGS space and embeds watermarks into diffusion-editing-preserved low-frequency Gaussians via coordinated covariance regularization and 2D filtering. In addition, we exploit the low-pass filtering behavior of diffusion-based editing by using Gaussian blur as an efficient training surrogate, enabling adversarial fine-tuning that further enhances watermark robustness against diffusion-based editing. Empirically, comprehensive quantitative and qualitative evaluations on three benchmark datasets demonstrate that RDSplat not only maintains superior robustness under diffusion-based editing, but also preserves watermark invisibility, achieving state-of-the-art performance.

Large Language Models (LLMs) have demonstrated exceptional capabilities in natural language understanding and generation. Based on these LLMs, businesses have started to provide Embeddings-as-a-Service (EaaS), offering feature extraction capabilities (in the form of text embeddings) that benefit downstream natural language processing tasks. However, prior research has demonstrated that EaaS is vulnerable to imitation attacks, where an attacker clones the service's model in a black-box manner without access to the model's internal workings. In response, watermarks have been added to the text embeddings to protect the intellectual property of EaaS providers by allowing them to check for model ownership. This thesis focuses on defending against imitation attacks by investigating EaaS watermarks. To achieve this goal, we unveil novel attacks and propose and validate new watermarking techniques. Firstly, we show that existing EaaS watermarks can be removed through paraphrasing the input text when attackers clone the model during imitation attacks. Our study illustrates that paraphrasing can effectively bypass current state-of-the-art EaaS watermarks across various attack setups (including different paraphrasing techniques and models) and datasets in most instances. This demonstrates a new vulnerability in recent EaaS watermarking techniques. Subsequently, as a countermeasure, we propose a novel watermarking technique, WET (Watermarking EaaS with Linear Transformation), which employs linear transformation of the embeddings. Watermark verification is conducted by applying a reverse transformation and comparing the similarity between recovered and original embeddings. We demonstrate its robustness against paraphrasing attacks with near-perfect verifiability. We conduct detailed ablation studies to assess the significance of each component and hyperparameter in WET.

Backdoor attacks undermine the reliability and trustworthiness of machine learning systems by injecting hidden behaviors that can be maliciously activated at inference time. While such threats have been extensively studied in unimodal settings, their impact on multimodal foundation models, particularly vision-language models (VLMs), remains largely underexplored. In this work, we introduce \textbf{BackdoorVLM}, the first comprehensive benchmark for systematically evaluating backdoor attacks on VLMs across a broad range of settings. It adopts a unified perspective that injects and analyzes backdoors across core vision-language tasks, including image captioning and visual question answering. BackdoorVLM organizes multimodal backdoor threats into 5 representative categories: targeted refusal, malicious injection, jailbreak, concept substitution, and perceptual hijack. Each category captures a distinct pathway through which an adversary can manipulate a model's behavior. We evaluate these threats using 12 representative attack methods spanning text, image, and bimodal triggers, tested on 2 open-source VLMs and 3 multimodal datasets. Our analysis reveals that VLMs exhibit strong sensitivity to textual instructions, and in bimodal backdoors the text trigger typically overwhelms the image trigger when forming the backdoor mapping. Notably, backdoors involving the textual modality remain highly potent, with poisoning rates as low as 1\% yielding over 90\% success across most tasks. These findings highlight significant, previously underexplored vulnerabilities in current VLMs. We hope that BackdoorVLM can serve as a useful benchmark for analyzing and mitigating multimodal backdoor threats. Code is available at: https://github.com/bin015/BackdoorVLM .

Backdoor attacks pose a serious threat to the secure deployment of large language models (LLMs), enabling adversaries to implant hidden behaviors triggered by specific inputs. However, existing methods often rely on manually crafted triggers and static data pipelines, which are rigid, labor-intensive, and inadequate for systematically evaluating modern defense robustness. As AI agents become increasingly capable, there is a growing need for more rigorous, diverse, and scalable \textit{red-teaming frameworks} that can realistically simulate backdoor threats and assess model resilience under adversarial conditions. In this work, we introduce \textsc{AutoBackdoor}, a general framework for automating backdoor injection, encompassing trigger generation, poisoned data construction, and model fine-tuning via an autonomous agent-driven pipeline. Unlike prior approaches, AutoBackdoor uses a powerful language model agent to generate semantically coherent, context-aware trigger phrases, enabling scalable poisoning across arbitrary topics with minimal human effort. We evaluate AutoBackdoor under three realistic threat scenarios, including \textit{Bias Recommendation}, \textit{Hallucination Injection}, and \textit{Peer Review Manipulation}, to simulate a broad range of attacks. Experiments on both open-source and commercial models, including LLaMA-3, Mistral, Qwen, and GPT-4o, demonstrate that our method achieves over 90\% attack success with only a small number of poisoned samples. More importantly, we find that existing defenses often fail to mitigate these attacks, underscoring the need for more rigorous and adaptive evaluation techniques against agent-driven threats as explored in this work. All code, datasets, and experimental configurations will be merged into our primary repository at https://github.com/bboylyg/BackdoorLLM.

The increasing realism of generated images has raised significant concerns about their potential misuse, necessitating robust detection methods. Current approaches mainly rely on training binary classifiers, which depend heavily on the quantity and quality of available generated images. In this work, we propose a novel framework that exploits geometric differences between the data manifolds of natural and generated images. To exploit this difference, we employ a pair of functions engineered to yield consistent outputs for natural images but divergent outputs for generated ones, leveraging the property that their gradients reside in mutually orthogonal subspaces. This design enables a simple yet effective detection method: an image is identified as generated if a transformation along its data manifold induces a significant change in the loss value of a self-supervised model pre-trained on natural images. Further more, to address diminishing manifold disparities in advanced generative models, we leverage normalizing flows to amplify detectable differences by extruding generated images away from the natural image manifold. Extensive experiments demonstrate the efficacy of this method. Code is available at https://github.com/tmlr-group/ConV.

As Large Language Models (LLMs) become increasingly integrated into software development workflows, they also become prime targets for adversarial attacks. Among these, backdoor attacks are a significant threat, allowing attackers to manipulate model outputs through hidden triggers embedded in training data. Detecting such backdoors remains a challenge, and one promising approach is the use of Spectral Signature defense methods that identify poisoned data by analyzing feature representations through eigenvectors. While some prior works have explored Spectral Signatures for backdoor detection in neural networks, recent studies suggest that these methods may not be optimally effective for code models. In this paper, we revisit the applicability of Spectral Signature-based defenses in the context of backdoor attacks on code models. We systematically evaluate their effectiveness under various attack scenarios and defense configurations, analyzing their strengths and limitations. We found that the widely used setting of Spectral Signature in code backdoor detection is often suboptimal. Hence, we explored the impact of different settings of the key factors. We discovered a new proxy metric that can more accurately estimate the actual performance of Spectral Signature without model retraining after the defense.

Decentralized Federated Learning enables privacy-preserving collaborative training without centralized servers but remains vulnerable to Byzantine attacks. Existing defenses require exchanging high-dimensional model vectors with all neighbors each round, creating prohibitive costs at scale. We propose SketchGuard, which decouples Byzantine filtering from aggregation via sketch-based screening. SketchGuard compresses $d$-dimensional models to $k$-dimensional sketches ($k \ll d$) using Count Sketch, then fetches full models only from accepted neighbors, reducing communication complexity from $O(d|N_i|)$ to $O(k|N_i| + d|S_i|)$, where $|N_i|$ is the neighbor count and $|S_i| \le |N_i|$ is the accepted count. We prove convergence in strongly convex and non-convex settings, showing that approximation errors introduce only a $(1+O(ε))$ factor in the effective threshold. Experiments demonstrate SketchGuard maintains state-of-the-art robustness (mean TER deviation $\leq$0.5 percentage points) while reducing computation by up to 82% and communication by 50-70%.

Object hallucination in Multimodal Large Language Models (MLLMs) is a persistent failure mode that causes the model to perceive objects absent in the image. This weakness of MLLMs is currently studied using static benchmarks with fixed visual scenarios, which preempts the possibility of uncovering model-specific or unanticipated hallucination vulnerabilities. We introduce GHOST (Generating Hallucinations via Optimizing Stealth Tokens), a method designed to stress-test MLLMs by actively generating images that induce hallucination. GHOST is fully automatic and requires no human supervision or prior knowledge. It operates by optimizing in the image embedding space to mislead the model while keeping the target object absent, and then guiding a diffusion model conditioned on the embedding to generate natural-looking images. The resulting images remain visually natural and close to the original input, yet introduce subtle misleading cues that cause the model to hallucinate. We evaluate our method across a range of models, including reasoning models like GLM-4.1V-Thinking, and achieve a hallucination success rate exceeding 28%, compared to around 1% in prior data-driven discovery methods. We confirm that the generated images are both high-quality and object-free through quantitative metrics and human evaluation. Also, GHOST uncovers transferable vulnerabilities: images optimized for Qwen2.5-VL induce hallucinations in GPT-4o at a 66.5% rate. Finally, we show that fine-tuning on our images mitigates hallucination, positioning GHOST as both a diagnostic and corrective tool for building more reliable multimodal systems.

Model inversion attacks (MIAs) aim to reconstruct class-representative samples from trained models. Recent generative MIAs utilize generative adversarial networks to learn image priors that guide the inversion process, yielding reconstructions with high visual quality and strong fidelity to the private training data. To explore the reason behind their effectiveness, we begin by examining the gradients of inversion loss with respect to synthetic inputs, and find that these gradients are surprisingly noisy. Further analysis reveals that generative inversion implicitly denoises these gradients by projecting them onto the tangent space of the generator manifold, filtering out off-manifold components while preserving informative directions aligned with the manifold. Our empirical measurements show that, in models trained with standard supervision, loss gradients often exhibit large angular deviations from the data manifold, indicating poor alignment with class-relevant directions. This observation motivates our central hypothesis: models become more vulnerable to MIAs when their loss gradients align more closely with the generator manifold. We validate this hypothesis by designing a novel training objective that explicitly promotes such alignment. Building on this insight, we further introduce a training-free approach to enhance gradient-manifold alignment during inversion, leading to consistent improvements over state-of-the-art generative MIAs.

Intelligent software systems powered by Large Language Models (LLMs) are increasingly deployed in critical sectors, raising concerns about their safety during runtime. Through an industry-academic collaboration when deploying an LLM-powered virtual customer assistant, a critical software engineering challenge emerged: how to enhance a safer deployment of LLM-powered software systems at runtime? While LlamaGuard, the current state-of-the-art runtime guardrail, offers protection against unsafe inputs, our study reveals a Defense Success Rate (DSR) drop of 24% under obfuscation- and template-based jailbreak attacks. In this paper, we propose DecipherGuard, a novel framework that integrates a deciphering layer to counter obfuscation-based prompts and a low-rank adaptation mechanism to enhance guardrail effectiveness against template-based attacks. Empirical evaluation on over 22,000 prompts demonstrates that DecipherGuard improves DSR by 36% to 65% and Overall Guardrail Performance (OGP) by 20% to 50% compared to LlamaGuard and two other runtime guardrails. These results highlight the effectiveness of DecipherGuard in defending LLM-powered software systems against jailbreak attacks during runtime.