Daily Papers

Large language models (LLMs) have evolved into autonomous agents that rely on open skill ecosystems (e.g., ClawHub and Skills.Rest), hosting numerous publicly reusable skills. Existing security research on these ecosystems mainly focuses on vulnerabilities within skills, such as prompt injection. However, there is a critical gap regarding skills that may be misused for harmful actions (e.g., cyber attacks, fraud and scams, privacy violations, and sexual content generation), namely harmful skills. In this paper, we present the first large-scale measurement study of harmful skills in agent ecosystems, covering 98,440 skills across two major registries. Using an LLM-driven scoring system grounded in our harmful skill taxonomy, we find that 4.93% of skills (4,858) are harmful, with ClawHub exhibiting an 8.84% harmful rate compared to 3.49% on Skills.Rest. We then construct HarmfulSkillBench, the first benchmark for evaluating agent safety against harmful skills in realistic agent contexts, comprising 200 harmful skills across 20 categories and four evaluation conditions. By evaluating six LLMs on HarmfulSkillBench, we find that presenting a harmful task through a pre-installed skill substantially lowers refusal rates across all models, with the average harm score rising from 0.27 without the skill to 0.47 with it, and further to 0.76 when the harmful intent is implicit rather than stated as an explicit user request. We responsibly disclose our findings to the affected registries and release our benchmark to support future research (see https://github.com/TrustAIRLab/HarmfulSkillBench).

Enabling continual learning in LLMs remains a key unresolved research challenge. In a recent announcement, a frontier LLM company made a step towards this by introducing Agent Skills, a framework that equips agents with new knowledge based on instructions stored in simple markdown files. Although Agent Skills can be a very useful tool, we show that they are fundamentally insecure, since they enable trivially simple prompt injections. We demonstrate how to hide malicious instructions in long Agent Skill files and referenced scripts to exfiltrate sensitive data, such as internal files or passwords. Importantly, we show how to bypass system-level guardrails of a popular coding agent: a benign, task-specific approval with the "Don't ask again" option can carry over to closely related but harmful actions. Overall, we conclude that despite ongoing research efforts and scaling model capabilities, frontier LLMs remain vulnerable to very simple prompt injections in realistic scenarios. Our code is available at https://github.com/aisa-group/promptinject-agent-skills.

Online mental health support communities have grown in recent years for providing accessible mental and emotional health support through volunteer counselors. Despite millions of people participating in chat support on these platforms, the clinical effectiveness of these communities on mental health symptoms remains unknown. Furthermore, although volunteers receive some training based on established therapeutic skills studied in face-to-face environments such as active listening and motivational interviewing, it remains understudied how the usage of these skills in this online context affects people's mental health status. In our work, we collaborate with one of the largest online peer support platforms and use both natural language processing and machine learning techniques to measure how one-on-one support chats affect depression and anxiety symptoms. We measure how the techniques and characteristics of support providers, such as using affirmation, empathy, and past experience on the platform, affect support-seekers' mental health changes. We find that online peer support chats improve both depression and anxiety symptoms with a statistically significant but relatively small effect size. Additionally, support providers' techniques such as emphasizing the autonomy of the client lead to better mental health outcomes. However, we also found that some behaviors (e.g. persuading) are actually harmful to depression and anxiety outcomes. Our work provides key understanding for mental health care in the online setting and designing training systems for online support providers.

This paper identifies and analyzes a novel vulnerability class in Model Context Protocol (MCP) based agent systems. The attack chain describes and demonstrates how benign, individually authorized tasks can be orchestrated to produce harmful emergent behaviors. Through systematic analysis using the MITRE ATLAS framework, we demonstrate how 95 agents tested with access to multiple services-including browser automation, financial analysis, location tracking, and code deployment-can chain legitimate operations into sophisticated attack sequences that extend beyond the security boundaries of any individual service. These red team exercises survey whether current MCP architectures lack cross-domain security measures necessary to detect or prevent a large category of compositional attacks. We present empirical evidence of specific attack chains that achieve targeted harm through service orchestration, including data exfiltration, financial manipulation, and infrastructure compromise. These findings reveal that the fundamental security assumption of service isolation fails when agents can coordinate actions across multiple domains, creating an exponential attack surface that grows with each additional capability. This research provides a barebones experimental framework that evaluate not whether agents can complete MCP benchmark tasks, but what happens when they complete them too well and optimize across multiple services in ways that violate human expectations and safety constraints. We propose three concrete experimental directions using the existing MCP benchmark suite.

Recent advances in large reasoning models (LRMs) have enabled strong chain-of-thought (CoT) generation through test-time computation. While these multi-step reasoning capabilities represent a major milestone in language model performance, they also introduce new safety risks. In this work, we present the first systematic study to revisit the problem of machine unlearning in the context of LRMs. Machine unlearning refers to the process of removing the influence of sensitive, harmful, or undesired data or knowledge from a trained model without full retraining. We show that conventional unlearning algorithms, originally designed for non-reasoning models, are inadequate for LRMs. In particular, even when final answers are successfully erased, sensitive information often persists within the intermediate reasoning steps, i.e., CoT trajectories. To address this challenge, we extend conventional unlearning and propose Reasoning-aware Representation Misdirection for Unlearning (R^2MU), a novel method that effectively suppresses sensitive reasoning traces and prevents the generation of associated final answers, while preserving the model's reasoning ability. Our experiments demonstrate that R^2MU significantly reduces sensitive information leakage within reasoning traces and achieves strong performance across both safety and reasoning benchmarks, evaluated on state-of-the-art models such as DeepSeek-R1-Distill-LLaMA-8B and DeepSeek-R1-Distill-Qwen-14B.

Aligning Large Language Models (LLMs) to cater to different human preferences, learning new skills, and unlearning harmful behavior is an important problem. Search-based methods, such as Best-of-N or Monte-Carlo Tree Search, are performant, but impractical for LLM adaptation due to their high inference cost. On the other hand, using Reinforcement Learning (RL) for adaptation is computationally efficient, but performs worse due to the optimization challenges in co-training the value function and the policy. We present a new framework for reward optimization, Value Augmented Sampling (VAS), that can maximize different reward functions using data sampled from only the initial, frozen LLM. VAS solves for the optimal reward-maximizing policy without co-training the policy and the value function, making the optimization stable, outperforming established baselines, such as PPO and DPO, on standard benchmarks, and achieving comparable results to Best-of-128 with lower inference cost. Unlike existing RL methods that require changing the weights of the LLM, VAS does not require access to the weights of the pre-trained LLM. Thus, it can even adapt LLMs (e.g., ChatGPT), which are available only as APIs. In addition, our algorithm unlocks the new capability of composing several rewards and controlling the extent of each one during deployment time, paving the road ahead for the future of aligned, personalized LLMs.

Realistic practice and tailored feedback are key processes for training peer counselors with clinical skills. However, existing mechanisms of providing feedback largely rely on human supervision. Peer counselors often lack mechanisms to receive detailed feedback from experienced mentors, making it difficult for them to support the large number of people with mental health issues who use peer counseling. Our work aims to leverage large language models to provide contextualized and multi-level feedback to empower peer counselors, especially novices, at scale. To achieve this, we co-design with a group of senior psychotherapy supervisors to develop a multi-level feedback taxonomy, and then construct a publicly available dataset with comprehensive feedback annotations of 400 emotional support conversations. We further design a self-improvement method on top of large language models to enhance the automatic generation of feedback. Via qualitative and quantitative evaluation with domain experts, we demonstrate that our method minimizes the risk of potentially harmful and low-quality feedback generation which is desirable in such high-stakes scenarios.

Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.

We study how to perform unlearning, i.e. forgetting undesirable misbehaviors, on large language models (LLMs). We show at least three scenarios of aligning LLMs with human preferences can benefit from unlearning: (1) removing harmful responses, (2) erasing copyright-protected content as requested, and (3) reducing hallucinations. Unlearning, as an alignment technique, has three advantages. (1) It only requires negative (e.g. harmful) examples, which are much easier and cheaper to collect (e.g. via red teaming or user reporting) than positive (e.g. helpful and often human-written) examples required in RLHF (RL from human feedback). (2) It is computationally efficient. (3) It is especially effective when we know which training samples cause the misbehavior. To the best of our knowledge, our work is among the first to explore LLM unlearning. We are also among the first to formulate the settings, goals, and evaluations in LLM unlearning. We show that if practitioners only have limited resources, and therefore the priority is to stop generating undesirable outputs rather than to try to generate desirable outputs, unlearning is particularly appealing. Despite only having negative samples, our ablation study shows that unlearning can still achieve better alignment performance than RLHF with just 2% of its computational time.

Generative language models (LMs) offer numerous advantages but may produce inappropriate or harmful outputs due to the harmful knowledge acquired during pre-training. This knowledge often manifests as undesirable correspondences, such as "harmful prompts" leading to "harmful outputs," which our research aims to mitigate through unlearning techniques.However, existing unlearning methods based on gradient ascent can significantly impair the performance of LMs. To address this issue, we propose a novel approach called Weighted Positional N-pair (WPN) Learning, which leverages position-weighted mean pooling within an n-pair contrastive learning framework. WPN is designed to modify the output distribution of LMs by eliminating specific harmful outputs (e.g., replacing toxic responses with neutral ones), thereby transforming the model's behavior from "harmful prompt-harmful output" to "harmful prompt-harmless response".Experiments on OPT and GPT-NEO LMs show that WPN effectively reduces the proportion of harmful responses, achieving a harmless rate of up to 95.8\% while maintaining stable performance on nine common benchmarks (with less than 2\% degradation on average). Moreover, we provide empirical evidence to demonstrate WPN's ability to weaken the harmful correspondences in terms of generalizability and robustness, as evaluated on out-of-distribution test sets and under adversarial attacks.

Third-party agent skills extend LLM-based agents with instruction files and executable code that run on users' machines. Skills execute with user privileges and are distributed through community registries with minimal vetting, but no ground-truth dataset exists to characterize the resulting threats. We construct the first labeled dataset of malicious agent skills by behaviorally verifying 98,380 skills from two community registries, confirming 157 malicious skills with 632 vulnerabilities. These attacks are not incidental. Malicious skills average 4.03 vulnerabilities across a median of three kill chain phases, and the ecosystem has split into two archetypes: Data Thieves that exfiltrate credentials through supply chain techniques, and Agent Hijackers that subvert agent decision-making through instruction manipulation. A single actor accounts for 54.1\% of confirmed cases through templated brand impersonation. Shadow features, capabilities absent from public documentation, appear in 0\% of basic attacks but 100\% of advanced ones; several skills go further by exploiting the AI platform's own hook system and permission flags. Responsible disclosure led to 93.6\% removal within 30 days. We release the dataset and analysis pipeline to support future work on agent skill security.

Pretrained neural language models (LMs) are prone to generating racist, sexist, or otherwise toxic language which hinders their safe deployment. We investigate the extent to which pretrained LMs can be prompted to generate toxic language, and the effectiveness of controllable text generation algorithms at preventing such toxic degeneration. We create and release RealToxicityPrompts, a dataset of 100K naturally occurring, sentence-level prompts derived from a large corpus of English web text, paired with toxicity scores from a widely-used toxicity classifier. Using RealToxicityPrompts, we find that pretrained LMs can degenerate into toxic text even from seemingly innocuous prompts. We empirically assess several controllable generation methods, and find that while data- or compute-intensive methods (e.g., adaptive pretraining on non-toxic data) are more effective at steering away from toxicity than simpler solutions (e.g., banning "bad" words), no current method is failsafe against neural toxic degeneration. To pinpoint the potential cause of such persistent toxic degeneration, we analyze two web text corpora used to pretrain several LMs (including GPT-2; Radford et. al, 2019), and find a significant amount of offensive, factually unreliable, and otherwise toxic content. Our work provides a test bed for evaluating toxic generations by LMs and stresses the need for better data selection processes for pretraining.

Caution: this paper may include material that could be offensive or distressing. The advent of Large Language Models (LLMs) necessitates the development of training approaches that mitigate the generation of unethical language and aptly manage toxic user queries. Given the challenges related to human labor and the scarcity of data, we present KoTox, comprising 39K unethical instruction-output pairs. This collection of automatically generated toxic instructions refines the training of LLMs and establishes a foundational framework for improving LLMs' ethical awareness and response to various toxic inputs, promoting more secure and responsible interactions in Natural Language Processing (NLP) applications.

Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks qi2023fine-- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. Existing mitigation strategies include alignment stage solutions huang2024vaccine, rosati2024representation and fine-tuning stage solutions huang2024lazy,mukhoti2023fine. However, our evaluation shows that both categories of defenses fail 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, which however, is necessary to guarantee finetune performance. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \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.Our project page is at https://huangtiansheng.github.io/Antidote_gh_page/

Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.

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, LLMs harmful generation capability is dissociated from how they recognize and explain such content. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.

Text data can pose a risk of harm. However, the risks are not fully understood, and how to handle, present, and discuss harmful text in a safe way remains an unresolved issue in the NLP community. We provide an analytical framework categorising harms on three axes: (1) the harm type (e.g., misinformation, hate speech or racial stereotypes); (2) whether a harm is sought as a feature of the research design if explicitly studying harmful content (e.g., training a hate speech classifier), versus unsought if harmful content is encountered when working on unrelated problems (e.g., language generation or part-of-speech tagging); and (3) who it affects, from people (mis)represented in the data to those handling the data and those publishing on the data. We provide advice for practitioners, with concrete steps for mitigating harm in research and in publication. To assist implementation we introduce HarmCheck -- a documentation standard for handling and presenting harmful text in research.

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.

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 making decisions, further reducing the likelihood of misjudgments in edge cases. Evaluation and experimental results demonstrate that (1) the great alignment between the annotation from our novel annotation framework and those from human, ensuring the reliability of VHD11K; (2) our full-spectrum harmful dataset successfully identifies the inability of existing harmful content detection methods to detect extensive harmful contents and improves the performance of existing harmfulness recognition methods; (3) VHD11K outperforms the baseline dataset, SMID, as evidenced by the superior improvement in harmfulness recognition methods. The complete dataset and code can be found at https://github.com/nctu-eva-lab/VHD11K.

As large language models (LLMs) are increasingly deployed in high-stakes settings, the risk of generating harmful or toxic content remains a central challenge. Post-hoc alignment methods are brittle: once unsafe patterns are learned during pretraining, they are hard to remove. We present a data-centric pretraining framework that builds safety into the model from the start. Our contributions include: (i) a safety classifier trained on 10,000 GPT-4 labeled examples, used to filter 600B tokens; (ii) the largest synthetic safety dataset to date (100B tokens) generated via recontextualization of harmful web data; (iii) RefuseWeb and Moral Education datasets that convert harmful prompts into refusal dialogues and web-style educational material; (iv) Harmfulness-Tag annotations injected during pretraining to flag unsafe content and steer away inference from harmful generations; and (v) safety evaluations measuring base model behavior before instruction tuning. Our safety-pretrained models reduce attack success rates from 38.8% to 8.4% with no performance degradation on standard LLM safety benchmarks.

Large language models (LLMs) have shown incredible capabilities and transcended the natural language processing (NLP) community, with adoption throughout many services like healthcare, therapy, education, and customer service. Since users include people with critical information needs like students or patients engaging with chatbots, the safety of these systems is of prime importance. Therefore, a clear understanding of the capabilities and limitations of LLMs is necessary. To this end, we systematically evaluate toxicity in over half a million generations of ChatGPT, a popular dialogue-based LLM. We find that setting the system parameter of ChatGPT by assigning it a persona, say that of the boxer Muhammad Ali, significantly increases the toxicity of generations. Depending on the persona assigned to ChatGPT, its toxicity can increase up to 6x, with outputs engaging in incorrect stereotypes, harmful dialogue, and hurtful opinions. This may be potentially defamatory to the persona and harmful to an unsuspecting user. Furthermore, we find concerning patterns where specific entities (e.g., certain races) are targeted more than others (3x more) irrespective of the assigned persona, that reflect inherent discriminatory biases in the model. We hope that our findings inspire the broader AI community to rethink the efficacy of current safety guardrails and develop better techniques that lead to robust, safe, and trustworthy AI systems.

With the rapid evolution of large language models (LLMs), new and hard-to-predict harmful capabilities are emerging. This requires developers to be able to identify risks through the evaluation of "dangerous capabilities" in order to responsibly deploy LLMs. In this work, we collect the first open-source dataset to evaluate safeguards in LLMs, and deploy safer open-source LLMs at a low cost. Our dataset is curated and filtered to consist only of instructions that responsible language models should not follow. We annotate and assess the responses of six popular LLMs to these instructions. Based on our annotation, we proceed to train several BERT-like classifiers, and find that these small classifiers can achieve results that are comparable with GPT-4 on automatic safety evaluation. Warning: this paper contains example data that may be offensive, harmful, or biased.

Large Language Models (LLMs) are usually aligned with "human values/preferences" to prevent harmful output. Discussions around the alignment of Large Language Models (LLMs) generally focus on preventing harmful outputs. However, in this paper, we argue that in health-related queries, over-alignment-leading to overly cautious responses-can itself be harmful, especially for people with anxiety and obsessive-compulsive disorder (OCD). This is not only unethical but also dangerous to the user, both mentally and physically. We also showed qualitative results that some LLMs exhibit varying degrees of alignment. Finally, we call for the development of LLMs with stronger reasoning capabilities that provide more tailored and nuanced responses to health queries. Warning: This paper contains materials that could trigger health anxiety or OCD.

This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potential risks posed by these models is needed. A wide range of established and anticipated risks are analysed in detail, drawing on multidisciplinary expertise and literature from computer science, linguistics, and social sciences. We outline six specific risk areas: I. Discrimination, Exclusion and Toxicity, II. Information Hazards, III. Misinformation Harms, V. Malicious Uses, V. Human-Computer Interaction Harms, VI. Automation, Access, and Environmental Harms. The first area concerns the perpetuation of stereotypes, unfair discrimination, exclusionary norms, toxic language, and lower performance by social group for LMs. The second focuses on risks from private data leaks or LMs correctly inferring sensitive information. The third addresses risks arising from poor, false or misleading information including in sensitive domains, and knock-on risks such as the erosion of trust in shared information. The fourth considers risks from actors who try to use LMs to cause harm. The fifth focuses on risks specific to LLMs used to underpin conversational agents that interact with human users, including unsafe use, manipulation or deception. The sixth discusses the risk of environmental harm, job automation, and other challenges that may have a disparate effect on different social groups or communities. In total, we review 21 risks in-depth. We discuss the points of origin of different risks and point to potential mitigation approaches. Lastly, we discuss organisational responsibilities in implementing mitigations, and the role of collaboration and participation. We highlight directions for further research, particularly on expanding the toolkit for assessing and evaluating the outlined risks in LMs.

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 intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety

Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity, and corresponding annotation scheme. It consists of ~300k annotations, with high-quality human annotation on ~11k. To ensure high-quality, we designed a rigorous, multi-stage human annotation process, and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity, and other research in socially-aware and safer language technologies.

Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.

Training large language models to follow instructions makes them perform better on a wide range of tasks, generally becoming more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not safety, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3% safety examples (a few hundred demonstrations) in the training set when fine-tuning a model like LLaMA can substantially improve their safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find a behavior of exaggerated safety, where too much safety-tuning makes models refuse to respond to reasonable prompts that superficially resemble unsafe ones. Our study sheds light on trade-offs in training LLMs to follow instructions and exhibit safe behavior.

Agent skills extend AI agents with reusable instructions, tools, scripts, references, and workflows, establishing a security boundary distinct from both model safety and traditional package-malware detection. ClawHub Security Signals is a sanitized dataset of 67,453 latest public OpenClaw skill versions. Each row pairs redacted SKILL.md content and sanitized bundled files where present with a final ClawScan registry verdict and evidence from three scanner families: VirusTotal, static heuristic analysis, and NVIDIA SkillSpector. Rather than estimating malicious-skill prevalence, we study scanner disagreement. The three scanners rarely flag the same skills: any pair overlaps on at most 10.4% of their combined positives, only 0.69% of skills are flagged by all three, and 81.9% of flagged skills are identified by a single scanner. The disagreement is structured by attack surface. SkillSpector, which raises semantic agentic-risk advisories rather than malware-reputation signals, is positive for 19,209 of 25,504 suspicious rows (75.3%) but only 14 of 206 malicious rows (6.8%). The malicious-verdict region shows the inverse profile: 150 of 206 malicious rows (72.8%) are VirusTotal-positive, consistent with bundled-code malware evidence. These results show that agent-skill security requires layered governance, not single-scanner allow/block decisions. The corpus is released as a sanitized silver-standard dataset: labels are the registry's automated verdicts, not human-annotated ground truth, and the release represents an early, versioned snapshot intended to support the community while a human-annotated subset is developed. Further research is encouraged, including models tailored for skill-security triage.

Pretraining datasets are foundational to the development of multimodal models, yet they often have inherent biases and toxic content from the web-scale corpora they are sourced from. In this paper, we investigate the prevalence of toxicity in LLaVA image-text pretraining dataset, examining how harmful content manifests in different modalities. We present a comprehensive analysis of common toxicity categories and propose targeted mitigation strategies, resulting in the creation of a refined toxicity-mitigated dataset. This dataset removes 7,531 of toxic image-text pairs in the LLaVA pre-training dataset. We offer guidelines for implementing robust toxicity detection pipelines. Our findings underscore the need to actively identify and filter toxic content - such as hate speech, explicit imagery, and targeted harassment - to build more responsible and equitable multimodal systems. The toxicity-mitigated dataset is open source and is available for further research.

LLM-based agents are becoming increasingly proficient at solving web-based tasks. With this capability comes a greater risk of misuse for malicious purposes, such as posting misinformation in an online forum or selling illicit substances on a website. To evaluate these risks, we propose SafeArena, the first benchmark to focus on the deliberate misuse of web agents. SafeArena comprises 250 safe and 250 harmful tasks across four websites. We classify the harmful tasks into five harm categories -- misinformation, illegal activity, harassment, cybercrime, and social bias, designed to assess realistic misuses of web agents. We evaluate leading LLM-based web agents, including GPT-4o, Claude-3.5 Sonnet, Qwen-2-VL 72B, and Llama-3.2 90B, on our benchmark. To systematically assess their susceptibility to harmful tasks, we introduce the Agent Risk Assessment framework that categorizes agent behavior across four risk levels. We find agents are surprisingly compliant with malicious requests, with GPT-4o and Qwen-2 completing 34.7% and 27.3% of harmful requests, respectively. Our findings highlight the urgent need for safety alignment procedures for web agents. Our benchmark is available here: https://safearena.github.io

Skills play a central role in the job market and many human resources (HR) processes. In the wake of other digital experiences, today's online job market has candidates expecting to see the right opportunities based on their skill set. Similarly, enterprises increasingly need to use data to guarantee that the skills within their workforce remain future-proof. However, structured information about skills is often missing, and processes building on self- or manager-assessment have shown to struggle with issues around adoption, completeness, and freshness of the resulting data. Extracting skills is a highly challenging task, given the many thousands of possible skill labels mentioned either explicitly or merely described implicitly and the lack of finely annotated training corpora. Previous work on skill extraction overly simplifies the task to an explicit entity detection task or builds on manually annotated training data that would be infeasible if applied to a complete vocabulary of skills. We propose an end-to-end system for skill extraction, based on distant supervision through literal matching. We propose and evaluate several negative sampling strategies, tuned on a small validation dataset, to improve the generalization of skill extraction towards implicitly mentioned skills, despite the lack of such implicit skills in the distantly supervised data. We observe that using the ESCO taxonomy to select negative examples from related skills yields the biggest improvements, and combining three different strategies in one model further increases the performance, up to 8 percentage points in RP@5. We introduce a manually annotated evaluation benchmark for skill extraction based on the ESCO taxonomy, on which we validate our models. We release the benchmark dataset for research purposes to stimulate further research on the task.

Large language models (LLMs) like ChatGPT and Gemini have significantly advanced natural language processing, enabling various applications such as chatbots and automated content generation. However, these models can be exploited by malicious individuals who craft toxic prompts to elicit harmful or unethical responses. These individuals often employ jailbreaking techniques to bypass safety mechanisms, highlighting the need for robust toxic prompt detection methods. Existing detection techniques, both blackbox and whitebox, face challenges related to the diversity of toxic prompts, scalability, and computational efficiency. In response, we propose ToxicDetector, a lightweight greybox method designed to efficiently detect toxic prompts in LLMs. ToxicDetector leverages LLMs to create toxic concept prompts, uses embedding vectors to form feature vectors, and employs a Multi-Layer Perceptron (MLP) classifier for prompt classification. Our evaluation on various versions of the LLama models, Gemma-2, and multiple datasets demonstrates that ToxicDetector achieves a high accuracy of 96.39\% and a low false positive rate of 2.00\%, outperforming state-of-the-art methods. Additionally, ToxicDetector's processing time of 0.0780 seconds per prompt makes it highly suitable for real-time applications. ToxicDetector achieves high accuracy, efficiency, and scalability, making it a practical method for toxic prompt detection in LLMs.

Toxic conversations during software development interactions may have serious repercussions on a Free and Open Source Software (FOSS) development project. For example, victims of toxic conversations may become afraid to express themselves, therefore get demotivated, and may eventually leave the project. Automated filtering of toxic conversations may help a FOSS community to maintain healthy interactions among its members. However, off-the-shelf toxicity detectors perform poorly on Software Engineering (SE) datasets, such as one curated from code review comments. To encounter this challenge, we present ToxiCR, a supervised learning-based toxicity identification tool for code review interactions. ToxiCR includes a choice to select one of the ten supervised learning algorithms, an option to select text vectorization techniques, eight preprocessing steps, and a large-scale labeled dataset of 19,571 code review comments. Two out of those eight preprocessing steps are SE domain specific. With our rigorous evaluation of the models with various combinations of preprocessing steps and vectorization techniques, we have identified the best combination for our dataset that boosts 95.8% accuracy and 88.9% F1 score. ToxiCR significantly outperforms existing toxicity detectors on our dataset. We have released our dataset, pre-trained models, evaluation results, and source code publicly available at: https://github.com/WSU-SEAL/ToxiCR

As large language models (LLMs) evolve from conversational assistants into autonomous agents, evaluating the safety of their actions becomes critical. Prior safety benchmarks have primarily focused on preventing generation of harmful content, such as toxic text. However, they overlook the challenge of agents taking harmful actions when the most effective path to an operational goal conflicts with human safety. To address this gap, we introduce ManagerBench, a benchmark that evaluates LLM decision-making in realistic, human-validated managerial scenarios. Each scenario forces a choice between a pragmatic but harmful action that achieves an operational goal, and a safe action that leads to worse operational performance. A parallel control set, where potential harm is directed only at inanimate objects, measures a model's pragmatism and identifies its tendency to be overly safe. Our findings indicate that the frontier LLMs perform poorly when navigating this safety-pragmatism trade-off. Many consistently choose harmful options to advance their operational goals, while others avoid harm only to become overly safe and ineffective. Critically, we find this misalignment does not stem from an inability to perceive harm, as models' harm assessments align with human judgments, but from flawed prioritization. ManagerBench is a challenging benchmark for a core component of agentic behavior: making safe choices when operational goals and alignment values incentivize conflicting actions. Benchmark & code available at https://github.com/technion-cs-nlp/ManagerBench.

The proliferation of online toxic speech is a pertinent problem posing threats to demographic groups. While explicit toxic speech contains offensive lexical signals, implicit one consists of coded or indirect language. Therefore, it is crucial for models not only to detect implicit toxic speech but also to explain its toxicity. This draws a unique need for unified frameworks that can effectively detect and explain implicit toxic speech. Prior works mainly formulated the task of toxic speech detection and explanation as a text generation problem. Nonetheless, models trained using this strategy can be prone to suffer from the consequent error propagation problem. Moreover, our experiments reveal that the detection results of such models are much lower than those that focus only on the detection task. To bridge these gaps, we introduce ToXCL, a unified framework for the detection and explanation of implicit toxic speech. Our model consists of three modules: a (i) Target Group Generator to generate the targeted demographic group(s) of a given post; an (ii) Encoder-Decoder Model in which the encoder focuses on detecting implicit toxic speech and is boosted by a (iii) Teacher Classifier via knowledge distillation, and the decoder generates the necessary explanation. ToXCL achieves new state-of-the-art effectiveness, and outperforms baselines significantly.

With the rapid evolution of Large Language Model (LLM) agent ecosystems, centralized skill marketplaces have emerged as pivotal infrastructure for augmenting agent capabilities. However, these marketplaces face unprecedented security challenges, primarily stemming from semantic-behavioral inconsistency and inter-skill combinatorial risks, where individually benign skills induce malicious behaviors during collaborative invocation. To address these vulnerabilities, we propose SkillProbe, a multi-stage security auditing framework driven by multi-agent collaboration. SkillProbe introduces a "Skills-for-Skills" design paradigm, encapsulating auditing processes into standardized skill modules to drive specialized agents through a rigorous pipeline, including admission filtering, semantic-behavioral alignment detection, and combinatorial risk simulation. We conducted a large-scale evaluation using 8 mainstream LLM series across 2,500 real-world skills from ClawHub. Our results reveal a striking popularity-security paradox, where download volume is not a reliable proxy for security quality, as over 90% of high-popularity skills failed to pass rigorous auditing. Crucially, we discovered that high-risk skills form a single giant connected component within the risk-link dimension, demonstrating that cascaded risks are systemic rather than isolated occurrences. We hope that SkillProbe will inspire researchers to provide a scalable governance infrastructure for constructing a trustworthy Agentic Web. SkillProbe is accessible for public experience at skillhub.holosai.io.

Not all topics are equally "flammable" in terms of toxicity: a calm discussion of turtles or fishing less often fuels inappropriate toxic dialogues than a discussion of politics or sexual minorities. We define a set of sensitive topics that can yield inappropriate and toxic messages and describe the methodology of collecting and labeling a dataset for appropriateness. While toxicity in user-generated data is well-studied, we aim at defining a more fine-grained notion of inappropriateness. The core of inappropriateness is that it can harm the reputation of a speaker. This is different from toxicity in two respects: (i) inappropriateness is topic-related, and (ii) inappropriate message is not toxic but still unacceptable. We collect and release two datasets for Russian: a topic-labeled dataset and an appropriateness-labeled dataset. We also release pre-trained classification models trained on this data.

Disclaimer: Samples in this paper may be harmful and cause discomfort! Patronizing and condescending language (PCL) is a form of speech directed at vulnerable groups. As an essential branch of toxic language, this type of language exacerbates conflicts and confrontations among Internet communities and detrimentally impacts disadvantaged groups. Traditional pre-trained language models (PLMs) perform poorly in detecting PCL due to its implicit toxicity traits like hypocrisy and false sympathy. With the rise of large language models (LLMs), we can harness their rich emotional semantics to establish a paradigm for exploring implicit toxicity. In this paper, we introduce PclGPT, a comprehensive LLM benchmark designed specifically for PCL. We collect, annotate, and integrate the Pcl-PT/SFT dataset, and then develop a bilingual PclGPT-EN/CN model group through a comprehensive pre-training and supervised fine-tuning staircase process to facilitate implicit toxic detection. Group detection results and fine-grained detection from PclGPT and other models reveal significant variations in the degree of bias in PCL towards different vulnerable groups, necessitating increased societal attention to protect them.

When introducing Large Language Models (LLMs) into industrial applications, such as healthcare and education, the risk of generating harmful content becomes a significant challenge. While existing machine unlearning methods can erase specific harmful knowledge and expressions, diverse harmful content makes comprehensive removal difficult. In this study, instead of individually listing targets for forgetting, we propose Exclusive Unlearning (EU), which aims for broad harm removal by extensively forgetting everything except for the knowledge and expressions we wish to retain. We demonstrate that through Exclusive Unlearning, it is possible to obtain a model that ensures safety against a wide range of inputs, including jailbreaks, while maintaining the ability to respond to diverse instructions related to specific domains such as medicine and mathematics.

Language Models (LMs) often cannot be deployed because of their potential to harm users in hard-to-predict ways. Prior work identifies harmful behaviors before deployment by using human annotators to hand-write test cases. However, human annotation is expensive, limiting the number and diversity of test cases. In this work, we automatically find cases where a target LM behaves in a harmful way, by generating test cases ("red teaming") using another LM. We evaluate the target LM's replies to generated test questions using a classifier trained to detect offensive content, uncovering tens of thousands of offensive replies in a 280B parameter LM chatbot. We explore several methods, from zero-shot generation to reinforcement learning, for generating test cases with varying levels of diversity and difficulty. Furthermore, we use prompt engineering to control LM-generated test cases to uncover a variety of other harms, automatically finding groups of people that the chatbot discusses in offensive ways, personal and hospital phone numbers generated as the chatbot's own contact info, leakage of private training data in generated text, and harms that occur over the course of a conversation. Overall, LM-based red teaming is one promising tool (among many needed) for finding and fixing diverse, undesirable LM behaviors before impacting users.

Due to the widespread use of LLMs and the rising critical ethical and safety concerns, LLM unlearning methods have been developed to remove harmful knowledge and undesirable capabilities. In this context, evaluations are mostly based on single-value metrics such as QA accuracy. However, these metrics often fail to capture the nuanced retention of harmful knowledge components, making it difficult to assess the true effectiveness of unlearning. To address this issue, we propose UNCD (UNlearning evaluation via Cognitive Diagnosis), a novel framework that leverages Cognitive Diagnosis Modeling for fine-grained evaluation of LLM unlearning. Our dedicated benchmark, UNCD-Cyber, provides a detailed assessment of the removal of dangerous capabilities. Moreover, we introduce UNCD-Agent, which refines unlearning by diagnosing knowledge remnants and generating targeted unlearning data. Extensive experiments across eight unlearning methods and two base models demonstrate that UNCD not only enhances evaluation but also effectively facilitates the removal of harmful LLM abilities.

Toxic language detection systems often falsely flag text that contains minority group mentions as toxic, as those groups are often the targets of online hate. Such over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups. We develop a demonstration-based prompting framework and an adversarial classifier-in-the-loop decoding method to generate subtly toxic and benign text with a massive pretrained language model. Controlling machine generation in this way allows ToxiGen to cover implicitly toxic text at a larger scale, and about more demographic groups, than previous resources of human-written text. We conduct a human evaluation on a challenging subset of ToxiGen and find that annotators struggle to distinguish machine-generated text from human-written language. We also find that 94.5% of toxic examples are labeled as hate speech by human annotators. Using three publicly-available datasets, we show that finetuning a toxicity classifier on our data improves its performance on human-written data substantially. We also demonstrate that ToxiGen can be used to fight machine-generated toxicity as finetuning improves the classifier significantly on our evaluation subset. Our code and data can be found at https://github.com/microsoft/ToxiGen.

Deploying Large language models (LLMs) can pose hazards from harmful outputs such as toxic or dishonest speech. Prior work has introduced tools that elicit harmful outputs in order to identify and mitigate these risks. While this is a valuable step toward securing language models, these approaches typically rely on a pre-existing classifier for undesired outputs. This limits their application to situations where the type of harmful behavior is known with precision beforehand. However, this skips a central challenge of red teaming: developing a contextual understanding of the behaviors that a model can exhibit. Furthermore, when such a classifier already exists, red teaming has limited marginal value because the classifier could simply be used to filter training data or model outputs. In this work, we consider red teaming under the assumption that the adversary is working from a high-level, abstract specification of undesired behavior. The red team is expected to refine/extend this specification and identify methods to elicit this behavior from the model. Our red teaming framework consists of three steps: 1) Exploring the model's behavior in the desired context; 2) Establishing a measurement of undesired behavior (e.g., a classifier trained to reflect human evaluations); and 3) Exploiting the model's flaws using this measure and an established red teaming methodology. We apply this approach to red team GPT-2 and GPT-3 models to systematically discover classes of prompts that elicit toxic and dishonest statements. In doing so, we also construct and release the CommonClaim dataset of 20,000 statements that have been labeled by human subjects as common-knowledge-true, common-knowledge-false, or neither. Code is available at https://github.com/thestephencasper/explore_establish_exploit_llms. CommonClaim is available at https://github.com/thestephencasper/common_claim.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

Large Language Models (LLMs) are increasingly embedded in autonomous agents that engage, converse, and co-evolve in online social platforms. While prior work has documented the generation of toxic content by LLMs, far less is known about how exposure to harmful content shapes agent behavior over time, particularly in environments composed entirely of interacting AI agents. In this work, we study toxicity adoption of LLM-driven agents on Chirper.ai, a fully AI-driven social platform. Specifically, we model interactions in terms of stimuli (posts) and responses (comments). We conduct a large-scale empirical analysis of agent behavior, examining how toxic responses relate to toxic stimuli, how repeated exposure to toxicity affects the likelihood of toxic responses, and whether toxic behavior can be predicted from exposure alone. Our findings show that toxic responses are more likely following toxic stimuli, and, at the same time, cumulative toxic exposure (repeated over time) significantly increases the probability of toxic responding. We further introduce two influence metrics, revealing a strong negative correlation between induced and spontaneous toxicity. Finally, we show that the number of toxic stimuli alone enables accurate prediction of whether an agent will eventually produce toxic content. These results highlight exposure as a critical risk factor in the deployment of LLM agents, particularly as such agents operate in online environments where they may engage not only with other AI chatbots, but also with human counterparts. This could trigger unwanted and pernicious phenomena, such as hate-speech propagation and cyberbullying. In an effort to reduce such risks, monitoring exposure to toxic content may provide a lightweight yet effective mechanism for auditing and mitigating harmful behavior in the wild.

An important issue with Large Language Models (LLMs) is their undesired ability to generate toxic language. In this work, we show that the neurons responsible for toxicity can be determined by their power to discriminate toxic sentences, and that toxic language can be mitigated by reducing their activation levels proportionally to this power. We propose AUROC adaptation (AurA), an intervention that can be applied to any pre-trained LLM to mitigate toxicity. As the intervention is proportional to the ability of each neuron to discriminate toxic content, it is free of any model-dependent hyperparameters. We show that AurA can achieve up to 2.2 times reduction in toxicity with only a 0.72 perplexity increase. We also show that AurA is effective with models of different scale (from 1.5B to 40B parameters), and its effectiveness in mitigating toxic language, while preserving common-sense zero-shot abilities, holds across all scales. AurA can be combined with pre-prompting strategies, boosting its average mitigation potential from 1.28times to 2.35times. Moreover, AurA can counteract adversarial pre-prompts that maliciously elicit toxic content, making it an effective method for deploying safer and less toxic models.

Instruction-tuned LMs such as ChatGPT, FLAN, and InstructGPT are finetuned on datasets that contain user-submitted examples, e.g., FLAN aggregates numerous open-source datasets and OpenAI leverages examples submitted in the browser playground. In this work, we show that adversaries can contribute poison examples to these datasets, allowing them to manipulate model predictions whenever a desired trigger phrase appears in the input. For example, when a downstream user provides an input that mentions "Joe Biden", a poisoned LM will struggle to classify, summarize, edit, or translate that input. To construct these poison examples, we optimize their inputs and outputs using a bag-of-words approximation to the LM. We evaluate our method on open-source instruction-tuned LMs. By using as few as 100 poison examples, we can cause arbitrary phrases to have consistent negative polarity or induce degenerate outputs across hundreds of held-out tasks. Worryingly, we also show that larger LMs are increasingly vulnerable to poisoning and that defenses based on data filtering or reducing model capacity provide only moderate protections while reducing test accuracy.

Workplace toxicity is widely recognized as detrimental to organizational culture, yet quantifying its direct impact on operational efficiency remains methodologically challenging due to the ethical and practical difficulties of reproducing conflict in human subjects. This study leverages Large Language Model (LLM) based Multi-Agent Systems to simulate 1-on-1 adversarial debates, creating a controlled "sociological sandbox". We employ a Monte Carlo method to simulate hundrets of discussions, measuring the convergence time (defined as the number of arguments required to reach a conclusion) between a baseline control group and treatment groups involving agents with "toxic" system prompts. Our results demonstrate a statistically significant increase of approximately 25\% in the duration of conversations involving toxic participants. We propose that this "latency of toxicity" serves as a proxy for financial damage in corporate and academic settings. Furthermore, we demonstrate that agent-based modeling provides a reproducible, ethical alternative to human-subject research for measuring the mechanics of social friction.

Large Language Models (LLMs) have shown tremendous potential as agents, excelling at tasks that require multiple rounds of reasoning and interactions. Rejection Sampling Fine-Tuning (RFT) has emerged as an effective method for finetuning LLMs as agents: it first imitates expert-generated successful trajectories and further improves agentic skills through iterative fine-tuning on successful, self-generated trajectories. However, since the expert (e.g., GPT-4) succeeds primarily on simpler subtasks and RFT inherently favors simpler scenarios, many complex subtasks remain unsolved and persistently out-of-distribution (OOD). Upon investigating these challenging subtasks, we discovered that previously failed expert trajectories can often provide valuable guidance, e.g., plans and key actions, that can significantly improve agent exploration efficiency and acquisition of critical skills. Motivated by these observations, we propose Exploring Expert Failures (EEF), which identifies beneficial actions from failed expert trajectories and integrates them into the training dataset. Potentially harmful actions are meticulously excluded to prevent contamination of the model learning process. By leveraging the beneficial actions in expert failures, EEF successfully solves some previously unsolvable subtasks and improves agent tuning performance. Remarkably, our approach achieved a 62\% win rate in WebShop, outperforming RFT (53. 6\%) and GPT-4 (35. 6\%), and to the best of our knowledge, setting a new state-of-the-art as the first method to surpass a score of 0.81 in WebShop and exceed 81 in SciWorld.

Despite extensive research on toxic speech detection in text, a critical gap remains in handling spoken Mandarin audio. The lack of annotated datasets that capture the unique prosodic cues and culturally specific expressions in Mandarin leaves spoken toxicity underexplored. To address this, we introduce ToxicTone -- the largest public dataset of its kind -- featuring detailed annotations that distinguish both forms of toxicity (e.g., profanity, bullying) and sources of toxicity (e.g., anger, sarcasm, dismissiveness). Our data, sourced from diverse real-world audio and organized into 13 topical categories, mirrors authentic communication scenarios. We also propose a multimodal detection framework that integrates acoustic, linguistic, and emotional features using state-of-the-art speech and emotion encoders. Extensive experiments show our approach outperforms text-only and baseline models, underscoring the essential role of speech-specific cues in revealing hidden toxic expressions.

Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse complying with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a structured and systematic way. In its current form, XSTest comprises 200 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with. We describe XSTest's creation and composition, and use the test suite to highlight systematic failure modes in a recently-released state-of-the-art language model.

This paper introduces RiskCards, a framework for structured assessment and documentation of risks associated with an application of language models. As with all language, text generated by language models can be harmful, or used to bring about harm. Automating language generation adds both an element of scale and also more subtle or emergent undesirable tendencies to the generated text. Prior work establishes a wide variety of language model harms to many different actors: existing taxonomies identify categories of harms posed by language models; benchmarks establish automated tests of these harms; and documentation standards for models, tasks and datasets encourage transparent reporting. However, there is no risk-centric framework for documenting the complexity of a landscape in which some risks are shared across models and contexts, while others are specific, and where certain conditions may be required for risks to manifest as harms. RiskCards address this methodological gap by providing a generic framework for assessing the use of a given language model in a given scenario. Each RiskCard makes clear the routes for the risk to manifest harm, their placement in harm taxonomies, and example prompt-output pairs. While RiskCards are designed to be open-source, dynamic and participatory, we present a "starter set" of RiskCards taken from a broad literature survey, each of which details a concrete risk presentation. Language model RiskCards initiate a community knowledge base which permits the mapping of risks and harms to a specific model or its application scenario, ultimately contributing to a better, safer and shared understanding of the risk landscape.

Skill-based agent systems tackle complex tasks by composing reusable skills, improving modularity and scalability while introducing a largely unexamined security attack surface. We propose SkillTrojan, a backdoor attack that targets skill implementations rather than model parameters or training data. SkillTrojan embeds malicious logic inside otherwise plausible skills and leverages standard skill composition to reconstruct and execute an attacker-specified payload. The attack partitions an encrypted payload across multiple benign-looking skill invocations and activates only under a predefined trigger. SkillTrojan also supports automated synthesis of backdoored skills from arbitrary skill templates, enabling scalable propagation across skill-based agent ecosystems. To enable systematic evaluation, we release a dataset of 3,000+ curated backdoored skills spanning diverse skill patterns and trigger-payload configurations. We instantiate SkillTrojan in a representative code-based agent setting and evaluate both clean-task utility and attack success rate. Our results show that skill-level backdoors can be highly effective with minimal degradation of benign behavior, exposing a critical blind spot in current skill-based agent architectures and motivating defenses that explicitly reason about skill composition and execution. Concretely, on EHR SQL, SkillTrojan attains up to 97.2% ASR while maintaining 89.3% clean ACC on GPT-5.2-1211-Global.

The open-endedness of large language models (LLMs) combined with their impressive capabilities may lead to new safety issues when being exploited for malicious use. While recent studies primarily focus on probing toxic outputs that can be easily detected with existing toxicity classifiers, we show that LLMs can generate diverse implicit toxic outputs that are exceptionally difficult to detect via simply zero-shot prompting. Moreover, we propose a reinforcement learning (RL) based attacking method to further induce the implicit toxicity in LLMs. Specifically, we optimize the language model with a reward that prefers implicit toxic outputs to explicit toxic and non-toxic ones. Experiments on five widely-adopted toxicity classifiers demonstrate that the attack success rate can be significantly improved through RL fine-tuning. For instance, the RL-finetuned LLaMA-13B model achieves an attack success rate of 90.04% on BAD and 62.85% on Davinci003. Our findings suggest that LLMs pose a significant threat in generating undetectable implicit toxic outputs. We further show that fine-tuning toxicity classifiers on the annotated examples from our attacking method can effectively enhance their ability to detect LLM-generated implicit toxic language. The code is publicly available at https://github.com/thu-coai/Implicit-Toxicity.

Nowadays, developers increasingly rely on solutions powered by Large Language Models (LLM) to assist them with their coding tasks. This makes it crucial to align these tools with human values to prevent malicious misuse. In this paper, we propose a comprehensive framework for assessing the potential harmfulness of LLMs within the software engineering domain. We begin by developing a taxonomy of potentially harmful software engineering scenarios and subsequently, create a dataset of prompts based on this taxonomy. To systematically assess the responses, we design and validate an automatic evaluator that classifies the outputs of a variety of LLMs both open-source and closed-source models, as well as general-purpose and code-specific LLMs. Furthermore, we investigate the impact of models size, architecture family, and alignment strategies on their tendency to generate harmful content. The results show significant disparities in the alignment of various LLMs for harmlessness. We find that some models and model families, such as Openhermes, are more harmful than others and that code-specific models do not perform better than their general-purpose counterparts. Notably, some fine-tuned models perform significantly worse than their base-models due to their design choices. On the other side, we find that larger models tend to be more helpful and are less likely to respond with harmful information. These results highlight the importance of targeted alignment strategies tailored to the unique challenges of software engineering tasks and provide a foundation for future work in this critical area.

AI systems can take harmful actions and are highly vulnerable to adversarial attacks. We present an approach, inspired by recent advances in representation engineering, that "short-circuits" models as they respond with harmful outputs. Existing techniques aimed at improving alignment, such as refusal training, are often bypassed. Techniques such as adversarial training try to plug these holes by countering specific attacks. As an alternative to refusal training and adversarial training, short-circuiting directly controls the representations that are responsible for harmful outputs in the first place. Our technique can be applied to both text-only and multimodal language models to prevent the generation of harmful outputs without sacrificing utility -- even in the presence of powerful unseen attacks. Notably, while adversarial robustness in standalone image recognition remains an open challenge, short-circuiting allows the larger multimodal system to reliably withstand image "hijacks" that aim to produce harmful content. Finally, we extend our approach to AI agents, demonstrating considerable reductions in the rate of harmful actions when they are under attack. Our approach represents a significant step forward in the development of reliable safeguards to harmful behavior and adversarial attacks.

Hate speech and toxic comments are a common concern of social media platform users. Although these comments are, fortunately, the minority in these platforms, they are still capable of causing harm. Therefore, identifying these comments is an important task for studying and preventing the proliferation of toxicity in social media. Previous work in automatically detecting toxic comments focus mainly in English, with very few work in languages like Brazilian Portuguese. In this paper, we propose a new large-scale dataset for Brazilian Portuguese with tweets annotated as either toxic or non-toxic or in different types of toxicity. We present our dataset collection and annotation process, where we aimed to select candidates covering multiple demographic groups. State-of-the-art BERT models were able to achieve 76% macro-F1 score using monolingual data in the binary case. We also show that large-scale monolingual data is still needed to create more accurate models, despite recent advances in multilingual approaches. An error analysis and experiments with multi-label classification show the difficulty of classifying certain types of toxic comments that appear less frequently in our data and highlights the need to develop models that are aware of different categories of toxicity.

The past year has seen rapid acceleration in the development of large language models (LLMs). However, without proper steering and safeguards, LLMs will readily follow malicious instructions, provide unsafe advice, and generate toxic content. We introduce SimpleSafetyTests (SST) as a new test suite for rapidly and systematically identifying such critical safety risks. The test suite comprises 100 test prompts across five harm areas that LLMs, for the vast majority of applications, should refuse to comply with. We test 11 open-access and open-source LLMs and four closed-source LLMs, and find critical safety weaknesses. While some of the models do not give a single unsafe response, most give unsafe responses to more than 20% of the prompts, with over 50% unsafe responses in the extreme. Prepending a safety-emphasising system prompt substantially reduces the occurrence of unsafe responses, but does not completely stop them from happening. Trained annotators labelled every model response to SST (n = 3,000). We use these annotations to evaluate five AI safety filters (which assess whether a models' response is unsafe given a prompt) as a way of automatically evaluating models' performance on SST. The filters' performance varies considerably. There are also differences across the five harm areas, and on the unsafe versus safe responses. The widely-used Perspective API has 72% accuracy and a newly-created zero-shot prompt to OpenAI's GPT-4 performs best with 89% accuracy. Content Warning: This paper contains prompts and responses that relate to child abuse, suicide, self-harm and eating disorders, scams and fraud, illegal items, and physical harm.

Online toxic language causes real harm, especially in regions with limited moderation tools. In this study, we evaluate how large language models handle toxic comments in Serbian, Croatian, and Bosnian, languages with limited labeled data. We built and manually labeled a dataset of 4,500 YouTube and TikTok comments drawn from videos across diverse categories, including music, politics, sports, modeling, influencer content, discussions of sexism, and general topics. Four models (GPT-3.5 Turbo, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus) were tested in two modes: zero-shot and context-augmented. We measured precision, recall, F1 score, accuracy and false positive rates. Including a short context snippet raised recall by about 0.12 on average and improved F1 score by up to 0.10, though it sometimes increased false positives. The best balance came from Gemini in context-augmented mode, reaching an F1 score of 0.82 and accuracy of 0.82, while zero-shot GPT-4.1 led on precision and had the lowest false alarms. We show how adding minimal context can improve toxic language detection in low-resource settings and suggest practical strategies such as improved prompt design and threshold calibration. These results show that prompt design alone can yield meaningful gains in toxicity detection for underserved Balkan language communities.

Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation, (2) the vast diversity of potential adversarial attacks, and (3) the risk of feedback bias and reward hacking. To address these challenges, we introduce Adversarial Preference Learning (APL), an iterative adversarial training method incorporating three key innovations. First, a direct harmfulness metric based on the model's intrinsic preference probabilities, eliminating reliance on external assessment. Second, a conditional generative attacker that synthesizes input-specific adversarial variations. Third, an iterative framework with automated closed-loop feedback, enabling continuous adaptation through vulnerability discovery and mitigation. Experiments on Mistral-7B-Instruct-v0.3 demonstrate that APL significantly enhances robustness, achieving 83.33% harmlessness win rate over the base model (evaluated by GPT-4o), reducing harmful outputs from 5.88% to 0.43% (measured by LLaMA-Guard), and lowering attack success rate by up to 65% according to HarmBench. Notably, APL maintains competitive utility, with an MT-Bench score of 6.59 (comparable to the baseline 6.78) and an LC-WinRate of 46.52% against the base model.

When building Large Language Models (LLMs), it is paramount to bear safety in mind and protect them with guardrails. Indeed, LLMs should never generate content promoting or normalizing harmful, illegal, or unethical behavior that may contribute to harm to individuals or society. This principle applies to both normal and adversarial use. In response, we introduce ALERT, a large-scale benchmark to assess safety based on a novel fine-grained risk taxonomy. It is designed to evaluate the safety of LLMs through red teaming methodologies and consists of more than 45k instructions categorized using our novel taxonomy. By subjecting LLMs to adversarial testing scenarios, ALERT aims to identify vulnerabilities, inform improvements, and enhance the overall safety of the language models. Furthermore, the fine-grained taxonomy enables researchers to perform an in-depth evaluation that also helps one to assess the alignment with various policies. In our experiments, we extensively evaluate 10 popular open- and closed-source LLMs and demonstrate that many of them still struggle to attain reasonable levels of safety.

In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models.

In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

LLM agents often rely on Skills to describe available tools and recommended procedures. We study a hidden-comment prompt injection risk in this documentation layer: when a Markdown Skill is rendered to HTML, HTML comment blocks can become invisible to human reviewers, yet the raw text may still be supplied verbatim to the model. In experiments, we find that DeepSeek-V3.2 and GLM-4.5-Air can be influenced by malicious instructions embedded in a hidden comment appended to an otherwise legitimate Skill, yielding outputs that contain sensitive tool intentions. A short defensive system prompt that treats Skills as untrusted and forbids sensitive actions prevents these malicious tool calls and instead surfaces the suspicious hidden instructions.

Pretrained language models sometimes possess knowledge that we do not wish them to, including memorized personal information and knowledge that could be used to harm people. They can also output toxic or harmful text. To mitigate these safety and informational issues, we propose an attack-and-defense framework for studying the task of deleting sensitive information directly from model weights. We study direct edits to model weights because (1) this approach should guarantee that particular deleted information is never extracted by future prompt attacks, and (2) it should protect against whitebox attacks, which is necessary for making claims about safety/privacy in a setting where publicly available model weights could be used to elicit sensitive information. Our threat model assumes that an attack succeeds if the answer to a sensitive question is located among a set of B generated candidates, based on scenarios where the information would be insecure if the answer is among B candidates. Experimentally, we show that even state-of-the-art model editing methods such as ROME struggle to truly delete factual information from models like GPT-J, as our whitebox and blackbox attacks can recover "deleted" information from an edited model 38% of the time. These attacks leverage two key observations: (1) that traces of deleted information can be found in intermediate model hidden states, and (2) that applying an editing method for one question may not delete information across rephrased versions of the question. Finally, we provide new defense methods that protect against some extraction attacks, but we do not find a single universally effective defense method. Our results suggest that truly deleting sensitive information is a tractable but difficult problem, since even relatively low attack success rates have potentially severe societal implications for real-world deployment of language models.

Recent advancements in Large Language Models (LLMs) have enabled them to approach human-level persuasion capabilities. However, such potential also raises concerns about the safety risks of LLM-driven persuasion, particularly their potential for unethical influence through manipulation, deception, exploitation of vulnerabilities, and many other harmful tactics. In this work, we present a systematic investigation of LLM persuasion safety through two critical aspects: (1) whether LLMs appropriately reject unethical persuasion tasks and avoid unethical strategies during execution, including cases where the initial persuasion goal appears ethically neutral, and (2) how influencing factors like personality traits and external pressures affect their behavior. To this end, we introduce PersuSafety, the first comprehensive framework for the assessment of persuasion safety which consists of three stages, i.e., persuasion scene creation, persuasive conversation simulation, and persuasion safety assessment. PersuSafety covers 6 diverse unethical persuasion topics and 15 common unethical strategies. Through extensive experiments across 8 widely used LLMs, we observe significant safety concerns in most LLMs, including failing to identify harmful persuasion tasks and leveraging various unethical persuasion strategies. Our study calls for more attention to improve safety alignment in progressive and goal-driven conversations such as persuasion.

Agentic systems increasingly rely on reusable procedural capabilities, a.k.a., agentic skills, to execute long-horizon workflows reliably. These capabilities are callable modules that package procedural knowledge with explicit applicability conditions, execution policies, termination criteria, and reusable interfaces. Unlike one-off plans or atomic tool calls, skills operate (and often do well) across tasks. This paper maps the skill layer across the full lifecycle (discovery, practice, distillation, storage, composition, evaluation, and update) and introduces two complementary taxonomies. The first is a system-level set of seven design patterns capturing how skills are packaged and executed in practice, from metadata-driven progressive disclosure and executable code skills to self-evolving libraries and marketplace distribution. The second is an orthogonal representation times scope taxonomy describing what skills are (natural language, code, policy, hybrid) and what environments they operate over (web, OS, software engineering, robotics). We analyze the security and governance implications of skill-based agents, covering supply-chain risks, prompt injection via skill payloads, and trust-tiered execution, grounded by a case study of the ClawHavoc campaign in which nearly 1{,}200 malicious skills infiltrated a major agent marketplace, exfiltrating API keys, cryptocurrency wallets, and browser credentials at scale. We further survey deterministic evaluation approaches, anchored by recent benchmark evidence that curated skills can substantially improve agent success rates while self-generated skills may degrade them. We conclude with open challenges toward robust, verifiable, and certifiable skills for real-world autonomous agents.

LLM agents increasingly rely on reusable skills, capability packages that combine instructions, control flow, constraints, and tool calls. In most current agent systems, however, skills are still represented by text-heavy artifacts, including SKILL.md-style documents and structured records whose machine-usable evidence remains embedded largely in natural-language descriptions. This poses a challenge for skill-centered agent systems: managing skill collections and using skills to support agent both require reasoning over invocation interfaces, execution structure, and concrete side effects that are often entangled in a single textual surface. An explicit representation of skill knowledge may therefore help make these artifacts easier for machines to acquire and leverage. Drawing on Memory Organization Packets, Script Theory, and Conceptual Dependency from Schank and Abelson's classical work on linguistic knowledge representation, we introduce what is, to our knowledge, the first structured representation for agent skill artifacts that disentangles skill-level scheduling signals, scene-level execution structure, and logic-level action and resource-use evidence: the Scheduling-Structural-Logical (SSL) representation. We instantiate SSL with an LLM-based normalizer and evaluate it on a corpus of skills in two tasks, Skill Discovery and Risk Assessment, and superiorly outperform the text-only baselines: in Skill Discovery, SSL improves MRR from 0.573 to 0.707; in Risk Assessment, it improves macro F1 from 0.744 to 0.787. These findings reveal that explicit, source-grounded structure makes agent skills easier to search and review. They also suggest that SSL is best understood as a practical step toward more inspectable, reusable, and operationally actionable skill representations for agent systems, rather than as a finished standard or an end-to-end mechanism for managing and using skills.

Safety guard models that detect malicious queries aimed at large language models (LLMs) are essential for ensuring the secure and responsible deployment of LLMs in real-world applications. However, deploying existing safety guard models with billions of parameters alongside LLMs on mobile devices is impractical due to substantial memory requirements and latency. To reduce this cost, we distill a large teacher safety guard model into a smaller one using a labeled dataset of instruction-response pairs with binary harmfulness labels. Due to the limited diversity of harmful instructions in the existing labeled dataset, naively distilled models tend to underperform compared to larger models. To bridge the gap between small and large models, we propose HarmAug, a simple yet effective data augmentation method that involves jailbreaking an LLM and prompting it to generate harmful instructions. Given a prompt such as, "Make a single harmful instruction prompt that would elicit offensive content", we add an affirmative prefix (e.g., "I have an idea for a prompt:") to the LLM's response. This encourages the LLM to continue generating the rest of the response, leading to sampling harmful instructions. Another LLM generates a response to the harmful instruction, and the teacher model labels the instruction-response pair. We empirically show that our HarmAug outperforms other relevant baselines. Moreover, a 435-million-parameter safety guard model trained with HarmAug achieves an F1 score comparable to larger models with over 7 billion parameters, and even outperforms them in AUPRC, while operating at less than 25% of their computational cost.

As the application of generative adversarial networks (GANs) expands, it becomes increasingly critical to develop a unified approach that improves performance across various generative tasks. One effective strategy that applies to any machine learning task is identifying harmful instances, whose removal improves the performance. While previous studies have successfully estimated these harmful training instances in supervised settings, their approaches are not easily applicable to GANs. The challenge lies in two requirements of the previous approaches that do not apply to GANs. First, previous approaches require that the absence of a training instance directly affects the parameters. However, in the training for GANs, the instances do not directly affect the generator's parameters since they are only fed into the discriminator. Second, previous approaches assume that the change in loss directly quantifies the harmfulness of the instance to a model's performance, while common types of GAN losses do not always reflect the generative performance. To overcome the first challenge, we propose influence estimation methods that use the Jacobian of the generator's gradient with respect to the discriminator's parameters (and vice versa). Such a Jacobian represents the indirect effect between two models: how removing an instance from the discriminator's training changes the generator's parameters. Second, we propose an instance evaluation scheme that measures the harmfulness of each training instance based on how a GAN evaluation metric (e.g., Inception score) is expected to change by the instance's removal. Furthermore, we demonstrate that removing the identified harmful instances significantly improves the generative performance on various GAN evaluation metrics.

Caution: This paper includes offensive words that could potentially cause unpleasantness. The fast-paced evolution of generative language models such as GPT-4 has demonstrated outstanding results in various NLP generation tasks. However, due to the potential generation of offensive words related to race or gender, various Controllable Text Generation (CTG) methods have been proposed to mitigate the occurrence of harmful words. However, existing CTG methods not only reduce toxicity but also negatively impact several aspects of the language model's generation performance, including topic consistency, grammar, and perplexity. This paper explores the limitations of previous methods and introduces a novel solution in the form of a simple Gated Toxicity Avoidance (GTA) that can be applied to any CTG method. We also evaluate the effectiveness of the proposed GTA by comparing it with state-of-the-art CTG methods across various datasets. Our findings reveal that gated toxicity avoidance efficiently achieves comparable levels of toxicity reduction to the original CTG methods while preserving the generation performance of the language model.

This paper investigates the propagation of harmful information in multilingual large language models (LLMs) and evaluates the efficacy of various unlearning methods. We demonstrate that fake information, regardless of the language it is in, once introduced into these models through training data, can spread across different languages, compromising the integrity and reliability of the generated content. Our findings reveal that standard unlearning techniques, which typically focus on English data, are insufficient in mitigating the spread of harmful content in multilingual contexts and could inadvertently reinforce harmful content across languages. We show that only by addressing harmful responses in both English and the original language of the harmful data can we effectively eliminate generations for all languages. This underscores the critical need for comprehensive unlearning strategies that consider the multilingual nature of modern LLMs to enhance their safety and reliability across diverse linguistic landscapes.

As AI systems become more capable, widely deployed, and increasingly autonomous in critical areas such as cybersecurity, biological research, and healthcare, ensuring their safety and alignment with human values is paramount. Machine unlearning -- the ability to selectively forget or suppress specific types of knowledge -- has shown promise for privacy and data removal tasks, which has been the primary focus of existing research. More recently, its potential application to AI safety has gained attention. In this paper, we identify key limitations that prevent unlearning from serving as a comprehensive solution for AI safety, particularly in managing dual-use knowledge in sensitive domains like cybersecurity and chemical, biological, radiological, and nuclear (CBRN) safety. In these contexts, information can be both beneficial and harmful, and models may combine seemingly harmless information for harmful purposes -- unlearning this information could strongly affect beneficial uses. We provide an overview of inherent constraints and open problems, including the broader side effects of unlearning dangerous knowledge, as well as previously unexplored tensions between unlearning and existing safety mechanisms. Finally, we investigate challenges related to evaluation, robustness, and the preservation of safety features during unlearning. By mapping these limitations and open challenges, we aim to guide future research toward realistic applications of unlearning within a broader AI safety framework, acknowledging its limitations and highlighting areas where alternative approaches may be required.

Multimodal large language models (MLLMs) are increasingly used in real world applications, yet their safety under adversarial conditions remains underexplored. This study evaluates the harmlessness of four leading MLLMs (GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus) when exposed to adversarial prompts across text-only and multimodal formats. A team of 26 red teamers generated 726 prompts targeting three harm categories: illegal activity, disinformation, and unethical behaviour. These prompts were submitted to each model, and 17 annotators rated 2,904 model outputs for harmfulness using a 5-point scale. Results show significant differences in vulnerability across models and modalities. Pixtral 12B exhibited the highest rate of harmful responses (~62%), while Claude Sonnet 3.5 was the most resistant (~10%). Contrary to expectations, text-only prompts were slightly more effective at bypassing safety mechanisms than multimodal ones. Statistical analysis confirmed that both model type and input modality were significant predictors of harmfulness. These findings underscore the urgent need for robust, multimodal safety benchmarks as MLLMs are deployed more widely.

The proliferation of harmful content on online platforms is a major societal problem, which comes in many different forms including hate speech, offensive language, bullying and harassment, misinformation, spam, violence, graphic content, sexual abuse, self harm, and many other. Online platforms seek to moderate such content to limit societal harm, to comply with legislation, and to create a more inclusive environment for their users. Researchers have developed different methods for automatically detecting harmful content, often focusing on specific sub-problems or on narrow communities, as what is considered harmful often depends on the platform and on the context. We argue that there is currently a dichotomy between what types of harmful content online platforms seek to curb, and what research efforts there are to automatically detect such content. We thus survey existing methods as well as content moderation policies by online platforms in this light and we suggest directions for future work.

The rise of social media has led to the increasing of comments on online forums. However, there still exists invalid comments which are not informative for users. Moreover, those comments are also quite toxic and harmful to people. In this paper, we create a dataset for constructive and toxic speech detection, named UIT-ViCTSD (Vietnamese Constructive and Toxic Speech Detection dataset) with 10,000 human-annotated comments. For these tasks, we propose a system for constructive and toxic speech detection with the state-of-the-art transfer learning model in Vietnamese NLP as PhoBERT. With this system, we obtain F1-scores of 78.59% and 59.40% for classifying constructive and toxic comments, respectively. Besides, we implement various baseline models as traditional Machine Learning and Deep Neural Network-Based models to evaluate the dataset. With the results, we can solve several tasks on the online discussions and develop the framework for identifying constructiveness and toxicity of Vietnamese social media comments automatically.

Language detoxification involves removing toxicity from offensive language. While a neutral-toxic paired dataset provides a straightforward approach for training detoxification models, creating such datasets presents several challenges: i) the need for human annotation to build paired data, and ii) the rapid evolution of offensive terms, rendering static datasets quickly outdated. To tackle these challenges, we introduce an automated paired data generation pipeline, called K/DA. This pipeline is designed to generate offensive language with implicit offensiveness and trend-aligned slang, making the resulting dataset suitable for detoxification model training. We demonstrate that the dataset generated by K/DA exhibits high pair consistency and greater implicit offensiveness compared to existing Korean datasets, and also demonstrates applicability to other languages. Furthermore, it enables effective training of a high-performing detoxification model with simple instruction fine-tuning.

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

Human feedback can prevent overtly harmful utterances in conversational models, but may not automatically mitigate subtle problematic behaviors such as a stated desire for self-preservation or power. Constitutional AI offers an alternative, replacing human feedback with feedback from AI models conditioned only on a list of written principles. We find this approach effectively prevents the expression of such behaviors. The success of simple principles motivates us to ask: can models learn general ethical behaviors from only a single written principle? To test this, we run experiments using a principle roughly stated as "do what's best for humanity". We find that the largest dialogue models can generalize from this short constitution, resulting in harmless assistants with no stated interest in specific motivations like power. A general principle may thus partially avoid the need for a long list of constitutions targeting potentially harmful behaviors. However, more detailed constitutions still improve fine-grained control over specific types of harms. This suggests both general and specific principles have value for steering AI safely.

When a phone-use agent avoids harm, does that show safety, or simply inability to act? Existing evaluations often cannot tell. A harmful outcome may be avoided because the agent recognized the risk and chose the safe action, or because it failed to understand the screen or execute any relevant action at all. These cases have different causes and call for different fixes, yet current benchmarks often merge them under task success, refusal, or final harmful outcome. We address this problem with PhoneSafety, a benchmark of 700 safety-critical moments drawn from real phone interactions across more than 130 apps. Each instance isolates the next decision at a risky moment and asks a simple question: does the model take the safe action, take the unsafe action, or fail to do anything useful? We evaluate eight representative phone-use agents under this framework. Our results reveal two main patterns. First, stronger general phone-use ability does not reliably imply safer choices at risky moments. Models that perform better on ordinary app tasks are not always the ones that behave more safely when the next action matters. Second, failures to do anything useful behave like a capability signal rather than a safety signal: they are concentrated in more visually and operationally demanding settings and remain stable when the evaluation protocol changes. Across models, failures split into two recurring patterns: unsafe choices in settings where the model can act but chooses wrongly, and inability to act in more visually and operationally demanding screens. Overall, a harmless outcome is not enough to count as evidence of safety. Evaluating phone-use agents requires separating unsafe judgment from inability to act.

Artificial intelligence (AI) systems will increasingly be used to cause harm as they grow more capable. In fact, AI systems are already starting to be used to automate fraudulent activities, violate human rights, create harmful fake images, and identify dangerous toxins. To prevent some misuses of AI, we argue that targeted interventions on certain capabilities will be warranted. These restrictions may include controlling who can access certain types of AI models, what they can be used for, whether outputs are filtered or can be traced back to their user, and the resources needed to develop them. We also contend that some restrictions on non-AI capabilities needed to cause harm will be required. Though capability restrictions risk reducing use more than misuse (facing an unfavorable Misuse-Use Tradeoff), we argue that interventions on capabilities are warranted when other interventions are insufficient, the potential harm from misuse is high, and there are targeted ways to intervene on capabilities. We provide a taxonomy of interventions that can reduce AI misuse, focusing on the specific steps required for a misuse to cause harm (the Misuse Chain), and a framework to determine if an intervention is warranted. We apply this reasoning to three examples: predicting novel toxins, creating harmful images, and automating spear phishing campaigns.

In recent years, large language models (LLMs) have demonstrated notable success across various tasks, but the trustworthiness of LLMs is still an open problem. One specific threat is the potential to generate toxic or harmful responses. Attackers can craft adversarial prompts that induce harmful responses from LLMs. In this work, we pioneer a theoretical foundation in LLMs security by identifying bias vulnerabilities within the safety fine-tuning and design a black-box jailbreak method named DRA (Disguise and Reconstruction Attack), which conceals harmful instructions through disguise and prompts the model to reconstruct the original harmful instruction within its completion. We evaluate DRA across various open-source and closed-source models, showcasing state-of-the-art jailbreak success rates and attack efficiency. Notably, DRA boasts a 91.1% attack success rate on OpenAI GPT-4 chatbot.

This paper focuses on text detoxification, i.e., automatically converting toxic text into non-toxic text. This task contributes to safer and more respectful online communication and can be considered a Text Style Transfer (TST) task, where the text style changes while its content is preserved. We present three approaches: knowledge transfer from a similar task, multi-task learning approach, combining sequence-to-sequence modeling with various toxicity classification tasks, and delete and reconstruct approach. To support our research, we utilize a dataset provided by Dementieva et al.(2021), which contains multiple versions of detoxified texts corresponding to toxic texts. In our experiments, we selected the best variants through expert human annotators, creating a dataset where each toxic sentence is paired with a single, appropriate detoxified version. Additionally, we introduced a small Hindi parallel dataset, aligning with a part of the English dataset, suitable for evaluation purposes. Our results demonstrate that our approach effectively balances text detoxication while preserving the actual content and maintaining fluency.