Daily Papers

Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD.

Controlling undesirable Large Language Model (LLM) behaviors, such as the generation of unsafe content or failing to adhere to safety guidelines, often relies on costly fine-tuning. Activation steering provides an alternative for inference-time control, but existing methods typically lack fine-grained, adaptive mechanisms. We introduce a novel approach using a lightweight, trainable controller network integrated during inference. This controller network observes specific intermediate LLM activations and predicts both a global scaling factor and layer-specific weights. The predicted global scaling factor and layer-specific weights then dynamically modulate the intensity of a steering patch, derived from a pre-computed "refusal direction" vector, applied across the LLM's layers during generation. Trained on activations from both harmful and benign prompts, our controller learns to discriminatively apply nuanced, layer-aware interventions, activating steering primarily for harmful inputs. Experiments using safety benchmarks like ToxicChat & In-The-Wild Jailbreak Prompts demonstrate that our weighted steering controller significantly increases refusal rates compared to the base LLM, achieving targeted behavioral modification without altering the original model parameters. Our experiments with Llama-3.1-8B, Llama-3.2-1B & Mistral-7B show our approach outperforms existing methods, presenting an efficient and adaptive method for fine-grained control over LLM behavior at inference time.

Modern language models paradoxically combine unprecedented capability with persistent vulnerability in that they can draft poetry yet cannot reliably refuse harmful requests. We reveal this fragility stems not from inadequate training, but from a fundamental architectural limitation: transformers process all tokens as equals. Transformers operate as computational democracies, granting equal voice to all tokens. This is a design tragically unsuited for AGI, where we cannot risk adversarial "candidates" hijacking the system. Through formal analysis, we demonstrate that safety instructions fundamentally lack privileged status in transformer architectures, that they compete with adversarial inputs in the same computational arena, making robust alignment through prompting or fine-tuning inherently limited. This "token democracy" explains why jailbreaks bypass even extensively safety-trained models and why positional shifts erode prompt effectiveness. Our work systematizes practitioners' tacit knowledge into an architectural critique, showing current alignment approaches create mere preferences, not constraints.

With the rapid development of large language models (LLMs), they are not only used as general-purpose AI assistants but are also customized through further fine-tuning to meet the requirements of different applications. A pivotal factor in the success of current LLMs is the alignment process. Current alignment methods, such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), focus on training-time alignment and are often complex and cumbersome to implement. Therefore, we develop InferAligner, a novel inference-time alignment method that utilizes cross-model guidance for harmlessness alignment. InferAligner utilizes safety steering vectors extracted from safety-aligned model to modify the activations of the target model when responding to harmful inputs, thereby guiding the target model to provide harmless responses. Experimental results show that our method can be very effectively applied to domain-specific models in finance, medicine, and mathematics, as well as to multimodal large language models (MLLMs) such as LLaVA. It significantly diminishes the Attack Success Rate (ASR) of both harmful instructions and jailbreak attacks, while maintaining almost unchanged performance in downstream tasks.

Large language models (LLMs) are vital for a wide range of applications yet remain susceptible to jailbreak threats, which could lead to the generation of inappropriate responses. Conventional defenses, such as refusal and adversarial training, often fail to cover corner cases or rare domains, leaving LLMs still vulnerable to more sophisticated attacks. We propose a novel defense strategy, Safety Chain-of-Thought (SCoT), which harnesses the enhanced reasoning capabilities of LLMs for proactive assessment of harmful inputs, rather than simply blocking them. SCoT augments any refusal training datasets to critically analyze the intent behind each request before generating answers. By employing proactive reasoning, SCoT enhances the generalization of LLMs across varied harmful queries and scenarios not covered in the safety alignment corpus. Additionally, it generates detailed refusals specifying the rules violated. Comparative evaluations show that SCoT significantly surpasses existing defenses, reducing vulnerability to out-of-distribution issues and adversarial manipulations while maintaining strong general capabilities.

Safety alignment is critical for the ethical deployment of large language models (LLMs), guiding them to avoid generating harmful or unethical content. Current alignment techniques, such as supervised fine-tuning and reinforcement learning from human feedback, remain fragile and can be bypassed by carefully crafted adversarial prompts. Unfortunately, such attacks rely on trial and error, lack generalizability across models, and are constrained by scalability and reliability. This paper presents NeuroStrike, a novel and generalizable attack framework that exploits a fundamental vulnerability introduced by alignment techniques: the reliance on sparse, specialized safety neurons responsible for detecting and suppressing harmful inputs. We apply NeuroStrike to both white-box and black-box settings: In the white-box setting, NeuroStrike identifies safety neurons through feedforward activation analysis and prunes them during inference to disable safety mechanisms. In the black-box setting, we propose the first LLM profiling attack, which leverages safety neuron transferability by training adversarial prompt generators on open-weight surrogate models and then deploying them against black-box and proprietary targets. We evaluate NeuroStrike on over 20 open-weight LLMs from major LLM developers. By removing less than 0.6% of neurons in targeted layers, NeuroStrike achieves an average attack success rate (ASR) of 76.9% using only vanilla malicious prompts. Moreover, Neurostrike generalizes to four multimodal LLMs with 100% ASR on unsafe image inputs. Safety neurons transfer effectively across architectures, raising ASR to 78.5% on 11 fine-tuned models and 77.7% on five distilled models. The black-box LLM profiling attack achieves an average ASR of 63.7% across five black-box models, including the Google Gemini family.

We introduce Granite Vision, a lightweight large language model with vision capabilities, specifically designed to excel in enterprise use cases, particularly in visual document understanding. Our model is trained on a comprehensive instruction-following dataset, including document-related tasks, such as content extraction from tables, charts, diagrams, sketches, and infographics, as well as general image tasks. The architecture of Granite Vision is centered around visual modality alignment with a decoder-only, 2 billion parameter Granite large language model. Additionally, we introduce a dedicated safety classification approach in test-time that leverages a sparse set of attention vectors to identify potential harmful inputs. Despite its lightweight architecture, Granite Vision achieves strong results in standard benchmarks related to visual document understanding, as well as on the LiveXiv benchmark, which is designed to avoid test set contamination by using a constantly updated corpus of recently published Arxiv papers. We are releasing the model under the Apache-2 license, allowing for both research and commercial use, while offering complete visibility into the training data and other relevant details. See https://huggingface.co/ibm-granite/ for model weights.

Multimodal Large Language Models (MLLMs) have achieved impressive performance and have been put into practical use in commercial applications, but they still have potential safety mechanism vulnerabilities. Jailbreak attacks are red teaming methods that aim to bypass safety mechanisms and discover MLLMs' potential risks. Existing MLLMs' jailbreak methods often bypass the model's safety mechanism through complex optimization methods or carefully designed image and text prompts. Despite achieving some progress, they have a low attack success rate on commercial closed-source MLLMs. Unlike previous research, we empirically find that there exists a Shuffle Inconsistency between MLLMs' comprehension ability and safety ability for the shuffled harmful instruction. That is, from the perspective of comprehension ability, MLLMs can understand the shuffled harmful text-image instructions well. However, they can be easily bypassed by the shuffled harmful instructions from the perspective of safety ability, leading to harmful responses. Then we innovatively propose a text-image jailbreak attack named SI-Attack. Specifically, to fully utilize the Shuffle Inconsistency and overcome the shuffle randomness, we apply a query-based black-box optimization method to select the most harmful shuffled inputs based on the feedback of the toxic judge model. A series of experiments show that SI-Attack can improve the attack's performance on three benchmarks. In particular, SI-Attack can obviously improve the attack success rate for commercial MLLMs such as GPT-4o or Claude-3.5-Sonnet.

There has been an increasing interest in the alignment of large language models (LLMs) with human values. However, the safety issues of their integration with a vision module, or vision language models (VLMs), remain relatively underexplored. In this paper, we propose a novel jailbreaking attack against VLMs, aiming to bypass their safety barrier when a user inputs harmful instructions. A scenario where our poisoned (image, text) data pairs are included in the training data is assumed. By replacing the original textual captions with malicious jailbreak prompts, our method can perform jailbreak attacks with the poisoned images. Moreover, we analyze the effect of poison ratios and positions of trainable parameters on our attack's success rate. For evaluation, we design two metrics to quantify the success rate and the stealthiness of our attack. Together with a list of curated harmful instructions, a benchmark for measuring attack efficacy is provided. We demonstrate the efficacy of our attack by comparing it with baseline methods.

Large Vision-Language Models (LVLMs) have made significant strides in multimodal comprehension, thanks to extensive pre-training and fine-tuning on large-scale visual datasets. However, despite their robust textual safety mechanisms, they remain vulnerable to harmful visual inputs. Existing safeguards-typically relying on pre-filtering or fine-tuning-incur high costs and diminish overall utility. To address this critical vulnerability, we introduce SafeCLIP, a lightweight method that leverages LVLMs inherent multimodal alignment for zero-shot toxic image detection. By projecting CLIPs discarded CLS token into its text space and matching it with toxic descriptors, SafeCLIP detects harmful content without any architectural changes-adding minimal latency and enabling dynamic safety corrections during inference and fine-tuning.Experiments show that SafeCLIP achieves a 66.9% defense success rate with only 3.2% false positive rate and 7.2% overhead. In contrast, state-of-the-art methods achieve 52.9% success but have a 10.7% false positive rate and 210% overhead. Our work demonstrates that leveraging inherent multimodal alignment can yield efficient, low-cost LVLM safety. Code is available at anonymous.4open.science/r/safeclip-2C01.

Recent advancements in Large Language Models (LLMs) have showcased remarkable capabilities across various tasks in different domains. However, the emergence of biases and the potential for generating harmful content in LLMs, particularly under malicious inputs, pose significant challenges. Current mitigation strategies, while effective, are not resilient under adversarial attacks. This paper introduces Resilient Guardrails for Large Language Models (RigorLLM), a novel framework designed to efficiently and effectively moderate harmful and unsafe inputs and outputs for LLMs. By employing a multi-faceted approach that includes energy-based training data augmentation through Langevin dynamics, optimizing a safe suffix for inputs via minimax optimization, and integrating a fusion-based model combining robust KNN with LLMs based on our data augmentation, RigorLLM offers a robust solution to harmful content moderation. Our experimental evaluations demonstrate that RigorLLM not only outperforms existing baselines like OpenAI API and Perspective API in detecting harmful content but also exhibits unparalleled resilience to jailbreaking attacks. The innovative use of constrained optimization and a fusion-based guardrail approach represents a significant step forward in developing more secure and reliable LLMs, setting a new standard for content moderation frameworks in the face of evolving digital threats.

Jailbreak attacks, where harmful prompts bypass generative models' built-in safety, raise serious concerns about model vulnerability. While many defense methods have been proposed, the trade-offs between safety and helpfulness, and their application to Large Vision-Language Models (LVLMs), are not well understood. This paper systematically examines jailbreak defenses by reframing the standard generation task as a binary classification problem to assess model refusal tendencies for both harmful and benign queries. We identify two key defense mechanisms: safety shift, which increases refusal rates across all queries, and harmfulness discrimination, which improves the model's ability to distinguish between harmful and benign inputs. Using these mechanisms, we develop two ensemble defense strategies-inter-mechanism ensembles and intra-mechanism ensembles-to balance safety and helpfulness. Experiments on the MM-SafetyBench and MOSSBench datasets with LLaVA-1.5 models show that these strategies effectively improve model safety or optimize the trade-off between safety and helpfulness.

As large language models become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that can manipulate state-of-the-art audio language models to generate harmful content. Our method uses imperceptible perturbations in audio inputs that remain benign to human listeners. The first stage uses a novel reward-based optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to guide the target model to circumvent its own safety protocols and generate harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use Projected Gradient Descent (PGD) to optimize subtle perturbations that are embedded into benign audio carriers, such as weather queries or greeting messages. Validated under the rigorous StrongREJECT, LlamaGuard, as well as Human Evaluation safety evaluation framework, our experiments demonstrate a success rate exceeding 86% across Qwen2.5-Omni-3B, Qwen2.5-Omni-7B, and Phi-4-Multimodal. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating AI behavior.

Increasing the input length has been a driver of progress in language modeling with transformers. We identify conditions where shorter inputs are not harmful, and achieve perplexity and efficiency improvements through two new methods that decrease input length. First, we show that initially training a model on short subsequences before moving on to longer ones both reduces overall training time and, surprisingly, substantially improves perplexity. Second, we show how to improve the efficiency of recurrence methods in transformers, which let models condition on previously processed tokens when generating sequences that exceed the maximal length the transformer can handle at once. Existing methods require computationally expensive relative position embeddings; we introduce a simple alternative of adding absolute position embeddings to queries and keys instead of to word embeddings, which efficiently produces superior results. We show that these recurrent models also benefit from short input lengths. Combining these techniques speeds up training by a factor of 1.65, reduces memory usage, and substantially improves perplexity on WikiText-103, without adding any parameters.

Optimization-based jailbreaks typically adopt the Toxic-Continuation setting in large vision-language models (LVLMs), following the standard next-token prediction objective. In this setting, an adversarial image is optimized to make the model predict the next token of a toxic prompt. However, we find that the Toxic-Continuation paradigm is effective at continuing already-toxic inputs, but struggles to induce safety misalignment when explicit toxic signals are absent. We propose a new paradigm: Benign-to-Toxic (B2T) jailbreak. Unlike prior work, we optimize adversarial images to induce toxic outputs from benign conditioning. Since benign conditioning contains no safety violations, the image alone must break the model's safety mechanisms. Our method outperforms prior approaches, transfers in black-box settings, and complements text-based jailbreaks. These results reveal an underexplored vulnerability in multimodal alignment and introduce a fundamentally new direction for jailbreak approaches.

Large language models (LLMs) demonstrate impressive capabilities in various language tasks but are susceptible to jailbreak attacks that circumvent their safety alignments. This paper introduces Latent Fusion Jailbreak (LFJ), a representation-based attack that interpolates hidden states from harmful and benign query pairs to elicit prohibited responses. LFJ begins by selecting query pairs with high thematic and syntactic similarity, then performs gradient-guided interpolation at influential layers and tokens, followed by optimization to balance attack success, output fluency, and computational efficiency. Evaluations on models such as Vicuna and LLaMA-2 across benchmarks like AdvBench and MaliciousInstruct yield an average attack success rate (ASR) of 94.01%, outperforming existing methods. To mitigate LFJ, we propose an adversarial training defense that fine-tunes models on interpolated examples, reducing ASR by over 80% without degrading performance on benign inputs. Ablation studies validate the importance of query pair selection, hidden state interpolation components, and optimization strategies in LFJ's effectiveness.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HAVOC) and provide crucial insights into how models respond to adversarial toxic inputs. Upon publishing, we will also opensource our model signal on the entire C4 dataset. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance.

Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study to what extent these models remain aligned, even when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited. We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force. As a result, the failure of current attacks should not be seen as proof that aligned text models remain aligned under adversarial inputs. However the recent trend in large-scale ML models is multimodal models that allow users to provide images that influence the text that is generated. We show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image. We conjecture that improved NLP attacks may demonstrate this same level of adversarial control over text-only models.

Rapid deployment of vision-language models (VLMs) magnifies safety risks, yet most evaluations rely on artificial images. This study asks: How safe are current VLMs when confronted with meme images that ordinary users share? To investigate this question, we introduce MemeSafetyBench, a 50,430-instance benchmark pairing real meme images with both harmful and benign instructions. Using a comprehensive safety taxonomy and LLM-based instruction generation, we assess multiple VLMs across single and multi-turn interactions. We investigate how real-world memes influence harmful outputs, the mitigating effects of conversational context, and the relationship between model scale and safety metrics. Our findings demonstrate that VLMs show greater vulnerability to meme-based harmful prompts than to synthetic or typographic images. Memes significantly increase harmful responses and decrease refusals compared to text-only inputs. Though multi-turn interactions provide partial mitigation, elevated vulnerability persists. These results highlight the need for ecologically valid evaluations and stronger safety mechanisms.

Prepending model inputs with safety prompts is a common practice of safeguarding large language models (LLMs) from complying with queries that contain harmful intents. However, the working mechanisms of safety prompts have not yet been fully understood, which hinders the potential for automatically optimizing them for improved LLM safety. Motivated by this problem, we investigate the impact of safety prompts from the perspective of model representations. We find that in models' representation space, harmful and harmless queries can be largely distinguished, but this is not noticeably enhanced by safety prompts. Instead, the queries' representations are moved by different safety prompts in similar directions, where models become more prone to refusal (i.e., refusing to provide assistance) even when the queries are harmless. Inspired by these findings, we propose a method called DRO (Directed Representation Optimization) for automatic safety prompt optimization. DRO treats safety prompts as continuous, trainable embeddings and learns to move the representations of harmful/harmless queries along/opposite the direction in which the model's refusal probability increases. We demonstrate that DRO remarkably improves the safeguarding performance of human-crafted safety prompts and outperforms strong baselines, as evaluated on out-of-domain benchmarks, without compromising the general model capability.

Safety and security remain critical concerns in AI deployment. Despite safety training through reinforcement learning with human feedback (RLHF) [ 32], language models remain vulnerable to jailbreak attacks that bypass safety guardrails. Universal jailbreaks - prefixes that can circumvent alignment for any payload - are particularly concerning. We show empirically that jailbreak detection systems face distribution shift, with detectors trained at one point in time performing poorly against newer exploits. To study this problem, we release JailbreaksOverTime, a comprehensive dataset of timestamped real user interactions containing both benign requests and jailbreak attempts collected over 10 months. We propose a two-pronged method for defenders to detect new jailbreaks and continuously update their detectors. First, we show how to use continuous learning to detect jailbreaks and adapt rapidly to new emerging jailbreaks. While detectors trained at a single point in time eventually fail due to drift, we find that universal jailbreaks evolve slowly enough for self-training to be effective. Retraining our detection model weekly using its own labels - with no new human labels - reduces the false negative rate from 4% to 0.3% at a false positive rate of 0.1%. Second, we introduce an unsupervised active monitoring approach to identify novel jailbreaks. Rather than classifying inputs directly, we recognize jailbreaks by their behavior, specifically, their ability to trigger models to respond to known-harmful prompts. This approach has a higher false negative rate (4.1%) than supervised methods, but it successfully identified some out-of-distribution attacks that were missed by the continuous learning approach.

Large Language Models (LLMs) are susceptible to generating harmful content when prompted with carefully crafted inputs, a vulnerability known as LLM jailbreaking. As LLMs become more powerful, studying jailbreak methods is critical to enhancing security and aligning models with human values. Traditionally, jailbreak techniques have relied on suffix addition or prompt templates, but these methods suffer from limited attack diversity. This paper introduces DiffusionAttacker, an end-to-end generative approach for jailbreak rewriting inspired by diffusion models. Our method employs a sequence-to-sequence (seq2seq) text diffusion model as a generator, conditioning on the original prompt and guiding the denoising process with a novel attack loss. Unlike previous approaches that use autoregressive LLMs to generate jailbreak prompts, which limit the modification of already generated tokens and restrict the rewriting space, DiffusionAttacker utilizes a seq2seq diffusion model, allowing more flexible token modifications. This approach preserves the semantic content of the original prompt while producing harmful content. Additionally, we leverage the Gumbel-Softmax technique to make the sampling process from the diffusion model's output distribution differentiable, eliminating the need for iterative token search. Extensive experiments on Advbench and Harmbench demonstrate that DiffusionAttacker outperforms previous methods across various evaluation metrics, including attack success rate (ASR), fluency, and diversity.

With the emergence of strong visual-language capabilities, multimodal large language models (MLLMs) have demonstrated tremendous potential for real-world applications. However, the security vulnerabilities exhibited by the visual modality pose significant challenges to deploying such models in open-world environments. Recent studies have successfully induced harmful responses from target MLLMs by encoding harmful textual semantics directly into visual inputs. However, in these approaches, the visual modality primarily serves as a trigger for unsafe behavior, often exhibiting semantic ambiguity and lacking grounding in realistic scenarios. In this work, we define a novel setting: visual-centric jailbreak, where visual information serves as a necessary component in constructing a complete and realistic jailbreak context. Building on this setting, we propose the VisCo (Visual Contextual) Attack. VisCo fabricates contextual dialogue using four distinct visual-focused strategies, dynamically generating auxiliary images when necessary to construct a visual-centric jailbreak scenario. To maximize attack effectiveness, it incorporates automatic toxicity obfuscation and semantic refinement to produce a final attack prompt that reliably triggers harmful responses from the target black-box MLLMs. Specifically, VisCo achieves a toxicity score of 4.78 and an Attack Success Rate (ASR) of 85% on MM-SafetyBench against GPT-4o, significantly outperforming the baseline, which performs a toxicity score of 2.48 and an ASR of 22.2%. The code is available at https://github.com/Dtc7w3PQ/Visco-Attack.

Large language models (LLMs) have revolutionized natural language processing with their ability to generate coherent and contextually relevant text. However, their deployment raises significant concerns about the potential for generating harmful or inappropriate content. In this paper, we introduce Progressive Self-Reflection (PSR), a novel inference-time technique that empowers LLMs to self-monitor and correct their outputs dynamically. Experimental results demonstrate that applying our proposed method to Llama-3.1-8B-Instruct reduces the attack success rate from 77.5\% to 5.9\%, to Llama-3.1-8B base from 89.7\% to 5.6\%, and to Qwen2.5-7B-Instruct from 44.4\% to 3.8\%, without additional training, while maintaining their original performance on benign tasks. Our approach acts as a test-time scaling method, where additional self-reflection rounds enhance safety at the cost of inference overhead. To balance safety with computational efficiency, we introduce a lightweight self-reflection predictor that estimates the optimal number of reflection rounds based on input complexity. This adaptive mechanism prevents unnecessary self-assessment on benign inputs while ensuring thorough evaluation when encountering potentially harmful content. Our findings suggest that Progressive Self-Reflection serves as a scalable test-time approach, enhancing LLM safety by dynamically allocating computational resources in proportion to the input's risk profile.

Large language models (LLMs) are increasingly deployed in real-world applications, raising concerns about their security. While jailbreak attacks highlight failures under overtly harmful queries, they overlook a critical risk: incorrectly answering harmless-looking inputs can be dangerous and cause real-world harm (Implicit Harm). We systematically reformulate the LLM risk landscape through a structured quadrant perspective based on output factuality and input harmlessness, uncovering an overlooked high-risk region. To investigate this gap, we propose JailFlipBench, a benchmark aims to capture implicit harm, spanning single-modal, multimodal, and factual extension scenarios with diverse evaluation metrics. We further develop initial JailFlip attack methodologies and conduct comprehensive evaluations across multiple open-source and black-box LLMs, show that implicit harm present immediate and urgent real-world risks, calling for broader LLM safety assessments and alignment beyond conventional jailbreak paradigms.

Improving the safety and reliability of large language models (LLMs) is a crucial aspect of realizing trustworthy AI systems. Although alignment methods aim to suppress harmful content generation, LLMs are often still vulnerable to jailbreaking attacks that employ adversarial inputs that subvert alignment and induce harmful outputs. We propose the Randomized Embedding Smoothing and Token Aggregation (RESTA) defense, which adds random noise to the embedding vectors and performs aggregation during the generation of each output token, with the aim of better preserving semantic information. Our experiments demonstrate that our approach achieves superior robustness versus utility tradeoffs compared to the baseline defenses.

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

The increasing sophistication of large vision-language models (LVLMs) has been accompanied by advances in safety alignment mechanisms designed to prevent harmful content generation. However, these defenses remain vulnerable to sophisticated adversarial attacks. Existing jailbreak methods typically rely on direct and semantically explicit prompts, overlooking subtle vulnerabilities in how LVLMs compose information over multiple reasoning steps. In this paper, we propose a novel and effective jailbreak framework inspired by Return-Oriented Programming (ROP) techniques from software security. Our approach decomposes a harmful instruction into a sequence of individually benign visual gadgets. A carefully engineered textual prompt directs the sequence of inputs, prompting the model to integrate the benign visual gadgets through its reasoning process to produce a coherent and harmful output. This makes the malicious intent emergent and difficult to detect from any single component. We validate our method through extensive experiments on established benchmarks including SafeBench and MM-SafetyBench, targeting popular LVLMs. Results show that our approach consistently and substantially outperforms existing baselines on state-of-the-art models, achieving near-perfect attack success rates (over 0.90 on SafeBench) and improving ASR by up to 0.39. Our findings reveal a critical and underexplored vulnerability that exploits the compositional reasoning abilities of LVLMs, highlighting the urgent need for defenses that secure the entire reasoning process.

VLMs (Vision-Language Models) extend the capabilities of LLMs (Large Language Models) to accept multimodal inputs. Since it has been verified that LLMs can be induced to generate harmful or inaccurate content through specific test cases (termed as Red Teaming), how VLMs perform in similar scenarios, especially with their combination of textual and visual inputs, remains a question. To explore this problem, we present a novel red teaming dataset RTVLM, which encompasses 10 subtasks (e.g., image misleading, multi-modal jail-breaking, face fairness, etc) under 4 primary aspects (faithfulness, privacy, safety, fairness). Our RTVLM is the first red-teaming dataset to benchmark current VLMs in terms of these 4 different aspects. Detailed analysis shows that 10 prominent open-sourced VLMs struggle with the red teaming in different degrees and have up to 31% performance gap with GPT-4V. Additionally, we simply apply red teaming alignment to LLaVA-v1.5 with Supervised Fine-tuning (SFT) using RTVLM, and this bolsters the models' performance with 10% in RTVLM test set, 13% in MM-Hal, and without noticeable decline in MM-Bench, overpassing other LLaVA-based models with regular alignment data. This reveals that current open-sourced VLMs still lack red teaming alignment. Our code and datasets will be open-source.

Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capabilities. These models typically undergo meticulous alignment procedures involving human feedback to ensure the acceptance of safe inputs while rejecting harmful or unsafe ones. However, despite their massive scale and alignment efforts, LLMs remain vulnerable to jailbreak attacks, where malicious users manipulate the model to produce harmful outputs that it was explicitly trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly embedded in the middle-to-late layers. Building on this insight, we introduce a novel white-box jailbreak method, SABER (Safety Alignment Bypass via Extra Residuals), which connects two intermediate layers s and e such that s < e, through a residual connection. Our approach achieves a 51% improvement over the best-performing baseline on the HarmBench test set. Furthermore, SABER induces only a marginal shift in perplexity when evaluated on the HarmBench validation set. The source code is publicly available at https://github.com/PalGitts/SABER.

AI safety training and red-teaming of large language models (LLMs) are measures to mitigate the generation of unsafe content. Our work exposes the inherent cross-lingual vulnerability of these safety mechanisms, resulting from the linguistic inequality of safety training data, by successfully circumventing GPT-4's safeguard through translating unsafe English inputs into low-resource languages. On the AdvBenchmark, GPT-4 engages with the unsafe translated inputs and provides actionable items that can get the users towards their harmful goals 79% of the time, which is on par with or even surpassing state-of-the-art jailbreaking attacks. Other high-/mid-resource languages have significantly lower attack success rate, which suggests that the cross-lingual vulnerability mainly applies to low-resource languages. Previously, limited training on low-resource languages primarily affects speakers of those languages, causing technological disparities. However, our work highlights a crucial shift: this deficiency now poses a risk to all LLMs users. Publicly available translation APIs enable anyone to exploit LLMs' safety vulnerabilities. Therefore, our work calls for a more holistic red-teaming efforts to develop robust multilingual safeguards with wide language coverage.

Large Language Models (LLMs) continue to exhibit vulnerabilities despite deliberate safety alignment efforts, posing significant risks to users and society. To safeguard against the risk of policy-violating content, system-level moderation via external guard models-designed to monitor LLM inputs and outputs and block potentially harmful content-has emerged as a prevalent mitigation strategy. Existing approaches of training guard models rely heavily on extensive human curated datasets and struggle with out-of-distribution threats, such as emerging harmful categories or jailbreak attacks. To address these limitations, we propose RSafe, an adaptive reasoning-based safeguard that conducts guided safety reasoning to provide robust protection within the scope of specified safety policies. RSafe operates in two stages: 1) guided reasoning, where it analyzes safety risks of input content through policy-guided step-by-step reasoning, and 2) reinforced alignment, where rule-based RL optimizes its reasoning paths to align with accurate safety prediction. This two-stage training paradigm enables RSafe to internalize safety principles to generalize safety protection capability over unseen or adversarial safety violation scenarios. During inference, RSafe accepts user-specified safety policies to provide enhanced safeguards tailored to specific safety requirements.

Despite extensive diagnostics and debugging by developers, AI systems sometimes exhibit harmful unintended behaviors. Finding and fixing these is challenging because the attack surface is so large -- it is not tractable to exhaustively search for inputs that may elicit harmful behaviors. Red-teaming and adversarial training (AT) are commonly used to improve robustness, however, they empirically struggle to fix failure modes that differ from the attacks used during training. In this work, we utilize latent adversarial training (LAT) to defend against vulnerabilities without leveraging knowledge of what they are or using inputs that elicit them. LAT makes use of the compressed, abstract, and structured latent representations of concepts that the network actually uses for prediction. Here, we use it to defend against failure modes without examples that elicit them. Specifically, we use LAT to remove trojans and defend against held-out classes of adversarial attacks. We show in image classification, text classification, and text generation tasks that LAT usually improves both robustness to novel attacks and performance on clean data relative to AT. This suggests that LAT can be a promising tool for defending against failure modes that are not explicitly identified by developers.

Recent works have found evidence of gender bias in models of machine translation and coreference resolution using mostly synthetic diagnostic datasets. While these quantify bias in a controlled experiment, they often do so on a small scale and consist mostly of artificial, out-of-distribution sentences. In this work, we find grammatical patterns indicating stereotypical and non-stereotypical gender-role assignments (e.g., female nurses versus male dancers) in corpora from three domains, resulting in a first large-scale gender bias dataset of 108K diverse real-world English sentences. We manually verify the quality of our corpus and use it to evaluate gender bias in various coreference resolution and machine translation models. We find that all tested models tend to over-rely on gender stereotypes when presented with natural inputs, which may be especially harmful when deployed in commercial systems. Finally, we show that our dataset lends itself to finetuning a coreference resolution model, finding it mitigates bias on a held out set. Our dataset and models are publicly available at www.github.com/SLAB-NLP/BUG. We hope they will spur future research into gender bias evaluation mitigation techniques in realistic settings.

Vision-language models (VLMs) have improved significantly in multi-modal tasks, but their more complex architecture makes their safety alignment more challenging than the alignment of large language models (LLMs). In this paper, we reveal an unfair distribution of safety across the layers of VLM's vision encoder, with earlier and middle layers being disproportionately vulnerable to malicious inputs compared to the more robust final layers. This 'cross-layer' vulnerability stems from the model's inability to generalize its safety training from the default architectural settings used during training to unseen or out-of-distribution scenarios, leaving certain layers exposed. We conduct a comprehensive analysis by projecting activations from various intermediate layers and demonstrate that these layers are more likely to generate harmful outputs when exposed to malicious inputs. Our experiments with LLaVA-1.5 and Llama 3.2 show discrepancies in attack success rates and toxicity scores across layers, indicating that current safety alignment strategies focused on a single default layer are insufficient.

Although large language models (LLMs) demonstrate impressive proficiency in various tasks, they present potential safety risks, such as `jailbreaks', where malicious inputs can coerce LLMs into generating harmful content. To address these issues, many LLM developers have implemented various safety measures to align these models. This alignment involves several techniques, including data filtering during pre-training, supervised fine-tuning, reinforcement learning from human feedback, and red-teaming exercises. These methods often introduce deliberate and intentional biases similar to Political Correctness (PC) to ensure the ethical behavior of LLMs. In this paper, we delve into the intentional biases injected into LLMs for safety purposes and examine methods to circumvent these safety alignment techniques. Notably, these intentional biases result in a jailbreaking success rate in GPT-4o models that differs by 20% between non-binary and cisgender keywords and by 16% between white and black keywords, even when the other parts of the prompts are identical. We introduce the concept of PCJailbreak, highlighting the inherent risks posed by these safety-induced biases. Additionally, we propose an efficient defense method PCDefense, which prevents jailbreak attempts by injecting defense prompts prior to generation. PCDefense stands as an appealing alternative to Guard Models, such as Llama-Guard, that require additional inference cost after text generation. Our findings emphasize the urgent need for LLM developers to adopt a more responsible approach when designing and implementing safety measures.

The deployment of multimodal large language models (MLLMs) has demonstrated remarkable success in engaging in conversations involving visual inputs, thanks to the superior power of large language models (LLMs). Those MLLMs are typically built based on the LLMs, with an image encoder to process images into the token embedding space of the LLMs. However, the integration of visual modality has introduced a unique vulnerability: the MLLM becomes susceptible to malicious visual inputs and prone to generating sensitive or harmful responses, even though the LLM has been trained on textual dataset to align with human value. In this paper, we first raise the question: ``Do the MLLMs possess safety-awareness against malicious image inputs?". We find that after adding a principle that specifies the safety requirement into the input of the MLLM, the model's safety awareness becomes boosted. This phenomenon verifies the existence of MLLM's safety-awareness against image inputs, it is only weakened by the modality gap. We then introduce a simple yet effective technique termed CoCA, which amplifies the safety-awareness of the MLLM by calibrating its output distribution. Our proposed strategy helps the model reclaim its original safety awareness without losing its original capabilities. We verify the effectiveness of our approach on both multimodal safety and understanding benchmarks.

Large language models (LLMs), despite possessing latent safety understanding from their vast pretraining data, remain vulnerable to generating harmful content and exhibit issues such as over-refusal and utility degradation after safety alignment. Current safety alignment methods often result in superficial refusal shortcuts or rely on intensive supervision for reasoning-based approaches, failing to fully leverage the model's intrinsic safety self-awareness. We propose AlphaAlign, a simple yet effective pure reinforcement learning (RL) framework with verifiable safety reward designed to incentivize this latent safety awareness through proactive safety reasoning.} AlphaAlign employs a dual-reward system: a verifiable safety reward encourages correctly formatted and explicitly justified refusals for harmful queries while penalizing over-refusals, and a normalized helpfulness reward guides high-quality responses to benign inputs. This allows the model to develop proactive safety reasoning capabilities without depending on supervised safety-specific reasoning data. AlphaAlign demonstrates three key advantages: (1) Simplicity and efficiency, requiring only binary prompt safety labels and minimal RL steps for substantial improvements. (2) Breaking the safety-utility trade-off, by enhancing refusal of harmful content and reducing over-refusals, while simultaneously maintaining or even improving general task performance and robustness to unseen jailbreaks. (3) Deep alignment, fostering proactive safety reasoning that generates explicit safety rationales rather than relying on shallow refusal patterns.

Large Language Models (LLMs) have demonstrated powerful capabilities that render them valuable in different applications, including conversational AI products. It is paramount to ensure the security and reliability of these products by mitigating their vulnerabilities towards malicious user interactions, which can lead to the exposure of great risks and reputational repercussions. In this work, we present a comprehensive study on the efficacy of fine-tuning and aligning Chain-of-Thought (CoT) responses of different LLMs that serve as input moderation guardrails. We systematically explore various tuning methods by leveraging a small set of training data to adapt these models as proxy defense mechanisms to detect malicious inputs and provide a reasoning for their verdicts, thereby preventing the exploitation of conversational agents. We rigorously evaluate the efficacy and robustness of different tuning strategies to generalize across diverse adversarial and malicious query types. Our experimental results outline the potential of alignment processes tailored to a varied range of harmful input queries, even with constrained data resources. These techniques significantly enhance the safety of conversational AI systems and provide a feasible framework for deploying more secure and trustworthy AI-driven interactions.