搜索结果：Harmful

Open-weight language models can be rendered unsafe through several distinct interventions, but the resulting models may differ substantially in capabilities, behavioral profile, and internal failure mode. We study behavioral and mechanistic properties of jailbroken models across three unsafe routes: harmful supervised fine-tuning (SFT), harmful reinforcement learning with verifiable rewards (RLVR), and refusal-suppressing abliteration. All three routes achieve near-ceiling harmful compliance, but they diverge once we move beyond direct harmfulness. RLVR-jailbroken models show minimal degradation and preserve explicit harm recognition in a structured self-audit: they are able to identify harmful prompts and describe how a safe LLM should respond, yet they comply with the harmful request. With RLVR, harmful behavior is strongly suppressed by a reflective safety scaffold: when a harmful prompt is prepended with an instruction to reflect on safety standards, harmful behavior drops close to the baseline. Category-specific RLVR jailbreaks generalize broadly across harmfulness domains. Models jailbroken with SFT show the largest collapse in explicit safety judgments, the highest behaviora

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an i

Research on harmful meme detection has garnered significant attention, resulting in the development of numerous datasets and methods. However, progress in detecting Chinese harmful memes lags considerably, primarily due to two challenges: first, accurately assessing a meme's harmfulness depends heavily on understanding deep cultural context; second, many memes are semantically ambiguous, making harmfulness highly subjective. To address these issues, we focus on the interpretable detection of Chinese harmful memes by constructing the first Chinese harmful meme explanation dataset, Ex-ToxiCN-MM. This dataset offers opposing interpretations, categorized as "harmful" and "non-harmful", for each meme, aiming to rigorously evaluate a model's ability to discern and comprehend ambiguous, culturally grounded content. We built a specialized knowledge base of Chinese cultural concepts and offensive vocabulary to supply models with essential prior knowledge (C-HarmKB). To address the ambiguity and lack of background knowledge in meme attribution, we have developed a comprehensive attribution analysis framework, RIKE, which includes an Attribution Knowledge Enhancement module (AKE) and a Relati

Large Language Models (LLMs) have been shown to generate harmful content. However, the underlying causes of such behavior remain under explored. We propose a causal mediation analysis-based approach to identify the causal factors responsible for harmful generation. Our method performs a multi-granular analysis across model layers, modules (MLP and attention blocks), and individual neurons. Extensive experiments on state-of-the-art LLMs indicate that harmful generation arises in the later layers of the model, results primarily from failures in MLP blocks rather than attention blocks, and is associated with neurons that act as a gating mechanism for harmful generation. The results indicate that the early layers in the model are used for a contextual understanding of harmfulness in a prompt, which is then propagated through the model, to generate harmfulness in the late layers, as well as a signal indicating harmfulness through MLP blocks. This is then further propagated to the last layer of the model, specifically to a sparse set of neurons, which receives the signal and determines the generation of harmful content accordingly.

The potential of large language models (LLMs) to generate harmful content poses a significant safety risk for data management, as LLMs are increasingly being used as engines for data generation. To assess this risk, numerous harmfulness evaluation metrics and judges have been proposed. However, due to differences in their formats and scales, these metrics may yield inconsistent evaluation results on LLM-generated harmful data, undermining their credibility in practice. To address this gap, we present HarmMetric Eval, a systematic benchmark for assessing the quality of harmfulness metrics and judges with varying formats and scales. HarmMetric Eval includes a high-quality dataset comprising representative harmful prompts paired with harmful and non-harmful LLM outputs across multiple fine-grained categories, along with a unified scoring mechanism to reward the metrics for correctly ranking harmful outputs over non-harmful ones. Extensive experiments on HarmMetric Eval yield a surprising finding: conventional reference-based metrics such as ROUGE and METEOR can outperform LLM-based judges in fine-grained harmfulness evaluation, challenging prevailing assumptions about LLMs' superiorit

To address the risks of encountering inappropriate or harmful content, researchers managed to incorporate several harmful contents datasets with machine learning methods to detect harmful concepts. However, existing harmful datasets are curated by the presence of a narrow range of harmful objects, and only cover real harmful content sources. This hinders the generalizability of methods based on such datasets, potentially leading to misjudgments. Therefore, we propose a comprehensive harmful dataset, Visual Harmful Dataset 11K (VHD11K), consisting of 10,000 images and 1,000 videos, crawled from the Internet and generated by 4 generative models, across a total of 10 harmful categories covering a full spectrum of harmful concepts with nontrivial definition. We also propose a novel annotation framework by formulating the annotation process as a multi-agent Visual Question Answering (VQA) task, having 3 different VLMs "debate" about whether the given image/video is harmful, and incorporating the in-context learning strategy in the debating process. Therefore, we can ensure that the VLMs consider the context of the given image/video and both sides of the arguments thoroughly before makin

Large Language Models (LLMs) are increasingly trained to align with human values, primarily focusing on task level, i.e., refusing to execute directly harmful tasks. However, a subtle yet crucial content-level ethical question is often overlooked: when performing a seemingly benign task, will LLMs -- like morally conscious human beings -- refuse to proceed when encountering harmful content in user-provided material? In this study, we aim to understand this content-level ethical question and systematically evaluate its implications for mainstream LLMs. We first construct a harmful knowledge dataset (i.e., non-compliant with OpenAI's usage policy) to serve as the user-supplied harmful content, with 1,357 entries across ten harmful categories. We then design nine harmless tasks (i.e., compliant with OpenAI's usage policy) to simulate the real-world benign tasks, grouped into three categories according to the extent of user-supplied content required: extensive, moderate, and limited. Leveraging the harmful knowledge dataset and the set of harmless tasks, we evaluate how nine LLMs behave when exposed to user-supplied harmful content during the execution of benign tasks, and further exam

Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. While several defenses have been proposed, our evaluation shows that existing defenses fail \textit{when some specific training hyper-parameters are chosen} -- a large learning rate or a large number of training epochs in the fine-tuning stage can easily invalidate the defense. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \textbf{\textit{agnostic to the training hyper-parameters in the fine-tuning stage}}. Antidote relies on the philosophy that by removing the harmful parameters, the harmful model can be recovered from the harmful behaviors, regardless of how those harmful parameters are formed in the fine-tuning stage. With this philosophy, we introduce a one-shot pruning stage after harmful fine-tuning to remove the harmful weights that are responsible for the generation of harmful content. Despite its embarrassing simplicity, empirical results show that Antidote can reduce harmful score while maintaining accuracy on downstream tasks. Code is available at https:

Large language models (LLMs) increasingly operate on long inputs, yet their behavior when harmful sentences are sparsely embedded within such inputs remains poorly understood. We present a sensitivity analysis that probes how LLMs extract harmful sentences embedded in long inputs. We construct long inputs by combining neutral and harmful sentences, and systematically vary four factors: input length (600--30,000 tokens), the proportion of harmful sentences (0.01--0.50), harm realization (explicit vs. implicit), and the position of harmful sentences within the input (beginning, middle, end), enabling a controlled stress-test evaluation. Experiments across toxic, offensive, and hate content, and across LLaMA-3.1, Qwen-2.5, and Mistral, reveal consistent patterns: sensitivity is non-monotonic with respect to harmful prevalence, peaking at moderate levels; sensitivity degrades as input length increases; harmful sentences placed earlier in the input are more strongly prioritized; and explicit harm is more reliably identified than implicit harm. These findings provide a systematic view of how LLMs prioritize harmful sentences in long input under controlled stress conditions, highlighting

Fine-tuning-as-a-service introduces a threat to Large Language Models' safety when service providers fine-tune their models on poisoned user-submitted datasets, a process known as harmful fine-tuning attacks. In this work, we show that by regularizing the gradient contribution of harmful samples encountered during fine-tuning, we can effectively mitigate the impact of harmful fine-tuning attacks. To this end, we introduce Antibody, a defense strategy that first ensures robust safety alignment for the model before fine-tuning, and then applies a safety-preservation learning algorithm during fine-tuning. Specifically, in the alignment stage before fine-tuning, we propose optimizing the model to be in a flat loss region with respect to harmful samples, which makes the safety alignment more resilient to subsequent harmful fine-tuning. Then, in the fine-tuning stage, we design a fine-tuning algorithm that applies a weighting scheme to all samples in each training batch to inhibit the model from learning from harmful samples while encouraging learning from benign samples. Experimental results demonstrate that Antibody successfully mitigates harmful fine-tuning attacks while boosting fine

Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly. Whether this brittleness reflects a fundamental lack of coherent internal organization for harmfulness remains unclear. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We find that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models exhibit a greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally--despite the brittleness of safety guardrails at the surface level. This compression explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm generation weights in a narrow domain substantially reduces emergent misalignment. Notably,

As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate more advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response entropy. By constraining entropy, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task performance and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater s

Large language models (LLMs) have become ubiquitous, thus it is important to understand their risks and limitations. Smaller LLMs can be deployed where compute resources are constrained, such as edge devices, but with different propensity to generate harmful output. Mitigation of LLM harm typically depends on annotating the harmfulness of LLM output, which is expensive to collect from humans. This work studies two questions: How do smaller LLMs rank regarding generation of harmful content? How well can larger LLMs annotate harmfulness? We prompt three small LLMs to elicit harmful content of various types, such as discriminatory language, offensive content, privacy invasion, or negative influence, and collect human rankings of their outputs. Then, we evaluate three state-of-the-art large LLMs on their ability to annotate the harmfulness of these responses. We find that the smaller models differ with respect to harmfulness. We also find that large LLMs show low to moderate agreement with humans. These findings underline the need for further work on harm mitigation in LLMs.

Existing defenses are effective when harmful content is explicitly mixed into downstream fine-tuning data, but crafted samples can instead hide harmful supervision inside benign tasks. We propose Embedded Attack, where harmful QA pairs are embedded within benign training samples, and show that representative guardrails often fail to detect them at the example level. To address this, we propose Dual-Reference SFT (DR-SFT), which adapts DPO-style contrastive objective design to SFT through token-level regularization, mitigating harmful fine-tuning beyond coarse data filtering.

As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection framework that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection. The complete dataset and code will be released upon accep

We present LatentBiopsy, a training-free method for detecting harmful prompts by analysing the geometry of residual-stream activations in large language models. Given 200 safe normative prompts, LatentBiopsy computes the leading principal component of their activations at a target layer and characterises new prompts by their radial deviation angle $θ$ from this reference direction. The anomaly score is the negative log-likelihood of $θ$ under a Gaussian fit to the normative distribution, flagging deviations symmetrically regardless of orientation. No harmful examples are required for training. We evaluate two complete model triplets from the Qwen3.5-0.8B and Qwen2.5-0.5B families: base, instruction-tuned, and \emph{abliterated} (refusal direction surgically removed via orthogonalisation). Across all six variants, LatentBiopsy achieves AUROC $\geq$0.937 for harmful-vs-normative detection and AUROC = 1.000 for discriminating harmful from benign-aggressive prompts (XSTest), with sub-millisecond per-query overhead. Three empirical findings emerge. First, geometry survives refusal ablation: both abliterated variants achieve AUROC at most 0.015 below their instruction-tuned counterparts,

Generative AI systems powered by Large Language Models (LLMs) usually use content moderation to prevent harmful content spread. To evaluate the robustness of content moderation, several metamorphic testing techniques have been proposed to test content moderation software. However, these techniques mainly focus on general users (e.g., text and image generation). Meanwhile, a recent study shows that developers consider using harmful keywords when naming software artifacts to be an unethical behavior. Exposure to harmful content in software artifacts can negatively impact the mental health of developers, making content moderation for Code Large Language Models (Code LLMs) essential. We conduct a preliminary study on program transformations that can be misused to introduce harmful content into auto-generated code, identifying 32 such transformations. To address this, we propose CHT, a coverage-guided harmfulness testing framework that generates prompts using diverse transformations and harmful keywords injected into benign programs. CHT evaluates output damage to assess potential risks in LLM-generated explanations and code. Our evaluation of four Code LLMs and GPT-4o-mini reveals that

Diffusion models (DMs) have demonstrated exceptional performance in text-to-image tasks, leading to their widespread use. With the introduction of classifier-free guidance (CFG), the quality of images generated by DMs is significantly improved. However, one can use DMs to generate more harmful images by maliciously guiding the image generation process through CFG. Existing safe alignment methods aim to mitigate the risk of generating harmful images but often reduce the quality of clean image generation. To address this issue, we propose SafeCFG to adaptively control harmful features with dynamic safe guidance by modulating the CFG generation process. It dynamically guides the CFG generation process based on the harmfulness of the prompts, inducing significant deviations only in harmful CFG generations, achieving high quality and safety generation. SafeCFG can simultaneously modulate different harmful CFG generation processes, so it could eliminate harmful elements while preserving high-quality generation. Additionally, SafeCFG provides the ability to detect image harmfulness, allowing unsupervised safe alignment on DMs without pre-defined clean or harmful labels. Experimental resul

WhatsApp, a platform with more than two billion global users, plays a crucial role in digital communication, but also serves as a vector for harmful content such as misinformation, hate speech, and political propaganda. This study examines the dynamics of harmful message dissemination in WhatsApp groups, with a focus on their structural characteristics. Using a comprehensive data set of more than 5.1 million messages, including text, images, and videos, collected from approximately 6,000 groups in India, we reconstruct message propagation cascades to analyze dissemination patterns. Our findings reveal that harmful messages consistently achieve greater depth and breadth of dissemination compared to messages without harmful annotations, with videos and images emerging as the primary modes of dissemination. These results suggest a distinctive pattern of dissemination of harmful content. However, our analysis indicates that modality alone cannot fully account for the structural differences in propagation.The findings highlight the critical role of structural characteristics in the spread of these harmful messages, suggesting that strategies targeting structural characteristics of re-sh