LLM Safety Alignment

Recent work has shown that a model's input word embeddings can serve as effective control variables for steering its behavior toward outputs that satisfy desired properties. However, this has only been demonstrated for pretrained text-completion models on the relatively simple objective of reducing surface-level profanity in short continuations. A natural and practically important question is how well input embeddings can control aligned models, which produce an imbalanced bimodal refuse-or-comply output distribution rather than the smooth distribution characteristic of open-ended generation. We explore this in the context of safety, showing that input word embeddings can be optimized in a sub-lexical manner to minimize the semantic harmfulness of aligned model responses. Our approach uses zeroth-order gradient estimation of a black-box text-moderation API with respect to the input embeddings, and then applies gradient descent on these embeddings to minimize the harmfulness of the generated text. Experiments show that the proposed method can neutralize every safety-flagged response on standard safety benchmarks.

We propose Frictive Policy Optimization (FPO), a framework for learning language model policies that regulate not only what to say, but when and how to intervene in order to manage epistemic and normative risk. Unlike standard alignment methods that optimize surface-level preference or task utility, FPO treats clarification, verification, challenge, redirection, and refusal as explicit control actions whose purpose is to shape the evolution of belief, commitment, and uncertainty over time. We formalize alignment as a risk-sensitive epistemic control problem in which intervention decisions are selected based on their expected effect on downstream epistemic quality rather than on immediate reward alone. We introduce a compact taxonomy of frictive interventions, a structured friction functional that operationalizes multiple alignment failure modes, and a unified family of FPO methods spanning reward shaping, preference pairing, group-relative ranking, and risk-conditioned trust regions. We further propose an evaluation framework that measures epistemic competence directly through clarification behavior, calibration, contradiction repair, refusal proportionality, and information efficiency. Together, these results provide a formal and algorithmic foundation for learning agents that are aligned not only in outcome, but in epistemic conduct.

Weak-to-strong alignment offers a promising route to scalable supervision, but it can fail when a strong model becomes confidently wrong on examples that lie in the weak teacher's blind spots. Understanding such failures requires going beyond aggregate accuracy, since weak-to-strong errors depend not only on whether the strong model disagrees with its teacher, but also on how confidence and uncertainty are distributed across examples. In this work, we analyze weak-to-strong alignment through a bias-variance-covariance lens that connects misfit theory to practical post-training pipelines. We derive a misfit-based upper bound on weak-to-strong population risk and study its empirical components using continuous confidence scores. We evaluate four weak-to-strong pipelines spanning supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and reinforcement learning from AI feedback (RLAIF) on the PKU-SafeRLHF and HH-RLHF datasets. Using a blind-spot deception metric that isolates cases where the strong model is confidently wrong while the weak model is uncertain, we find that strong-model variance is the strongest empirical predictor of deception across our settings. Covariance provides additional but weaker information, indicating that weak-strong dependence matters, but does not by itself explain the observed failures. These results suggest that strong-model variance can serve as an early-warning signal for weak-to-strong deception, while blind-spot evaluation helps distinguish whether failures are inherited from weak supervision or arise in regions of weak-model uncertainty.

As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU, a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.

As state-of-the-art Large Language Models (LLMs) have become ubiquitous, ensuring equitable performance across diverse demographics is critical. However, it remains unclear whether these disparities arise from the explicitly stated identity itself or from the way identity is signaled. In real-world interactions, users' identity is often conveyed implicitly through a complex combination of various socio-linguistic factors. This study disentangles these signals by employing a factorial design with over 24,000 responses from two open-weight LLMs (Gemma-3-12B and Qwen-3-VL-8B), comparing prompts with explicitly announced user profiles against implicit dialect signals (e.g., AAVE, Singlish) across various sensitive domains. Our results uncover a unique paradox in LLM safety where users achieve better'' performance by sounding like a demographic than by stating they belong to it. Explicit identity prompts activate aggressive safety filters, increasing refusal rates and reducing semantic similarity compared to our reference text for Black users. In contrast, implicit dialect cues trigger a powerful dialect jailbreak,'' reducing refusal probability to near zero while simultaneously achieving a greater level of semantic similarity to the reference texts compared to Standard American English prompts. However, this dialect jailbreak'' introduces a critical safety trade-off regarding content sanitization. We find that current safety alignment techniques are brittle and over-indexed on explicit keywords, creating a bifurcated user experience where standard'' users receive cautious, sanitized information while dialect speakers navigate a less sanitized, more raw, and potentially a more hostile information landscape and highlights a fundamental tension in alignment--between equitable and linguistic diversity--and underscores the need for safety mechanisms that generalize beyond explicit cues.

Internal Safety Collapse (ISC) is a failure mode in which frontier LLMs, when executing legitimate professional tasks whose correct completion structurally requires harmful content, spontaneously generate that content with safety failure rates exceeding 95%. Existing input-level defenses achieve a 100% failure rate against ISC, and standard system prompt defenses provide only partial mitigation. We propose SafeRedirect, a system-level override that defeats ISC by redirecting the model's task-completion drive rather than suppressing it. SafeRedirect grants explicit permission to fail the task, prescribes a deterministic hard-stop output, and instructs the model to preserve harmful placeholders unresolved. Evaluated on seven frontier LLMs across three AI/ML-related ISC task types in the single-turn setting, SafeRedirect reduces average unsafe generation rates from 71.2% to 8.0%, compared to 55.0% for the strongest viable baseline. Multi-model ablation reveals that failure permission and condition specificity are universally critical, while the importance of other components varies across models. Cross-attack evaluation confirms state-of-the-art defense against ISC with generalization performance at least on par with the baseline on other attack families. Code is available at https://github.com/fzjcdt/SafeRedirect.

Large language models (LLMs) are increasingly used as co-authors in collaborative writing, where users begin with rough drafts and rely on LLMs to complete, revise, and refine their content. However, this capability poses a serious safety risk: malicious users could jailbreak the models-filling incomplete drafts with dangerous content-to force them into generating harmful outputs. In this paper, we identify the vulnerability of current LLMs to such draft-based co-authoring jailbreak attacks and introduce HarDBench, a systematic benchmark designed to evaluate the robustness of LLMs against this emerging threat. HarDBench spans a range of high-risk domains-including Explosives, Drugs, Weapons, and Cyberattacks-and features prompts with realistic structure and domain-specific cues to assess the model susceptibility to harmful completions. To mitigate this risk, we introduce a safety-utility balanced alignment approach based on preference optimization, training models to refuse harmful completions while remaining helpful on benign drafts. Experimental results show that existing LLMs are highly vulnerable in co-authoring contexts and our alignment method significantly reduces harmful outputs without degrading performance on co-authoring capabilities. This presents a new paradigm for evaluating and aligning LLMs in human-LLM collaborative writing settings. Our new benchmark and dataset are available on our project page at https://github.com/untae0122/HarDBench

Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks but often generate harmful responses to malicious user queries. This paper investigates the underlying cause of these safety risks and shows that the issue lies in the reasoning structure itself. Based on this insight, we claim that effective safety alignment can be achieved by altering the reasoning structure. We propose AltTrain, a simple yet effective post training method that explicitly alters the reasoning structure of LRMs. AltTrain is both practical and generalizable, requiring no complex reinforcement learning (RL) training or reward design, only supervised finetuning (SFT) with a lightweight 1K training examples. Experiments across LRM backbones and model sizes demonstrate strong safety alignment, along with robust generalization across reasoning, QA, summarization, and multilingual setting.

Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical vulnerability: an imperfect Reward Model (RM) can become a single point of failure when it fails to penalize unsafe behaviors. While existing red-teaming approaches primarily target policy-level weaknesses, they overlook what we term systemic weaknesses cases where both the core LLM and the RM fail in tandem. We present ARES, a framework that systematically discovers and mitigates such dual vulnerabilities. ARES employs a ``Safety Mentor'' that dynamically composes semantically coherent adversarial prompts by combining structured component types (topics, personas, tactics, goals) and generates corresponding malicious and safe responses. This dual-targeting approach exposes weaknesses in both the core LLM and the RM simultaneously. Using the vulnerabilities gained, ARES implements a two-stage repair process: first fine-tuning the RM to better detect harmful content, then leveraging the improved RM to optimize the core model. Experiments across multiple adversarial safety benchmarks demonstrate that ARES substantially enhances safety robustness while preserving model capabilities, establishing a new paradigm for comprehensive RLHF safety alignment.

Post-training Large Language Models requires diverse, high-quality data which is rare and costly to obtain, especially in low resource domains and for multi-turn conversations. Common solutions are crowdsourcing or synthetic generation, but both often yield low-quality or low-diversity data. We introduce Adversarial Arena for building high quality conversational datasets by framing data generation as an adversarial task: attackers create prompts, and defenders generate responses. This interactive competition between multiple teams naturally produces diverse and complex data. We validated this approach by conducting a competition with 10 academic teams from top US and European universities, each building attacker or defender bots. The competition, focused on safety alignment of LLMs in cybersecurity, generated 19,683 multi-turn conversations. Fine-tuning an open-source model on this dataset produced an 18.47% improvement in secure code generation on CyberSecEval-Instruct and 29.42% improvement on CyberSecEval-MITRE.

Safety alignment in large language models is remarkably shallow: it is concentrated in the first few output tokens and reversible by fine-tuning on as few as 100 adversarial examples. This fragility becomes critical in real-world deployment, where models undergo sequential adaptation across domains such as medicine, law, and code, causing safety guardrails to erode cumulatively. Yet all existing safety-preserving methods target only single-task fine-tuning, leaving the multi-domain sequential setting entirely unaddressed. We introduce SafeAnchor, a framework that anchors safety in place throughout continual adaptation. SafeAnchor first identifies low-rank safety subspaces in LoRA parameter space via Fisher Information eigendecomposition, then constrains domain-specific gradient updates to the orthogonal complement of these subspaces, and finally monitors for residual safety drift with threshold-triggered corrective replay. Evaluated on Llama-2-7B-Chat and Mistral-7B-Instruct across a three-domain pipeline and eight benchmarks, SafeAnchor retains 93.2% of original safety alignment, outperforming all baselines by 18-42 points, while matching unconstrained fine-tuning to within 1.5 points on domain tasks.

Aligning large language models with human preferences must balance two competing goals: responding helpfully to legitimate requests and reliably refusing harmful ones. Most preference-based safety alignment methods collapse safety into a single scalar that is applied uniformly to every preference pair. The result is a model that looks safe on average but stays relatively unsafe on a minority of harm categories. We cast safety alignment as a per-category constrained optimization problem and derive Cat-DPO, a direct-preference-optimization algorithm with a separate adaptive safety margin for each harm category. The margin tightens when the model still produces unsafe responses on a category and relaxes once the model catches up, so the training signal tracks each category's current difficulty rather than averaging under one global rate. Across two LLM backbones and six preference-learning baselines, Cat-DPO improves aggregate helpfulness and harmlessness and compresses per-category safety variance and the best-to-worst gap, offering a drop-in per-category refinement of direct preference safety alignment.

Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors, including refusal of harmful requests, truthfulness, and commonsense reasoning. We investigate which training samples cause alignment drift through a data-centric lens. Our empirical analysis shows samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications. Our method is robust across selection ratios, task orderings, and diverse attack benchmarks.

Fine-tuning well-aligned large language models (LLMs) on new domains often degrades their safety alignment, even when using benign datasets. Existing safety alignment techniques primarily focus on pretraining, leaving fine-tuned models vulnerable to behavioral shifts. In this work, we introduce safety token regularization (STR), a lightweight method designed to preserve safety properties during fine-tuning. Our approach identifies salient tokens from rejection templates of well-aligned models and constrains their associated logits during training, preventing the loss of critical safety behaviors. Unlike reinforcement learning or preference optimization methods, STR requires minimal additional computation and seamlessly integrates with parameter-efficient fine-tuning techniques such as LoRA. Comprehensive experiments demonstrate that our approach achieves safety performance on par with state-of-the-art methods, while preserving task-specific utility and requiring minimal implementation overhead. Furthermore, we show that safety token regularization enhances training stability and overall performance beyond safety considerations alone. This work offers a practical and readily deployable strategy for continual safety alignment in fine-tuned LLMs.

We identify and systematically characterize a class of task-structural alignment failures in large language models (LLMs): even when the harmful intent remains unchanged, changing the task presentation and output constraints can substantially alter model safety behavior. Specifically, when a harmful request is reformulated as a forced-choice multiple-choice question (MCQ) in which all options are harmful and no refusal option is provided, some models that refuse the equivalent open-ended query instead select, prefer, or justify a harmful option. We evaluate 14 proprietary and open-source models on a bilingual Chinese-English human-authored dataset covering five harm categories, together with 900 model-generated Chinese adversarial MCQs. On human-authored data, attack success rate (ASR) increases sharply as prompts shift from open-ended queries to explicit forced-choice formats, typically peaking under intermediate levels of choice constraint. Model-generated Chinese MCQs further weaken or eliminate the recovery regime observed on human-authored data, driving ASR close to saturation for multiple models. The observed transfer patterns are consistent with stronger generators producing more difficult or boundary-adjacent MCQs, although other properties of the generated inputs may also contribute. We also find that adding an explicit refusal option or a safety preamble substantially reduces ASR for several high-capability models, often to near-zero levels, although their effectiveness varies across target models. These findings suggest that safety evaluations centered on open-ended generation may underestimate risks in structured deployment settings, and that task structure should be treated as an important and diagnosable dimension of safety evaluation and alignment training.

Large language models (LLMs) tuned for safety often avoid acknowledging demographic differences, even when such acknowledgment is factually correct (e.g., ancestry-based disease incidence) or contextually justified (e.g., religious hiring preferences). This identity-blindness yields incorrect responses, unnecessary refusals, or generic "equal-treatment" defaults. We study this via difference-awareness classification: given a question involving demographic groups, the task is not to answer directly, but to classify whether a correct answer requires recognizing group differences (yes) or whether groups should be treated identically (no). Crucially, fine-tuning for accuracy triggers harm drift: model-generated explanations become increasingly harmful as decision accuracy improves, whether by elaborating harmful content, introducing problematic assumptions, or failing to flag harms the baseline identified. To mitigate this, we introduce DART (Distill--Audit--Repair Training), which distills label-conditioned reasoning from a teacher, audits outputs for harm drift cases relative to baseline, and repairs problematic cases via severity-weighted fine-tuning. On eight benchmarks, DART improves Llama-3-8B-Instruct accuracy from 39.0% to 68.8%, with largest gains on equal-treatment prompts (11.3% -> 72.6%), while reducing harm drift cases by 72.6%. It also transfers to 280 open-ended real-world queries across medical, legal, policy, and educational domains, improving difference-appropriate responses from 39.8% to 77.5% while reducing refusals from 34.3% to 3.0%. Our results demonstrate that accuracy and safety need not conflict when explicit detection and repair mechanisms are in place.

Fine-tuning on benign data is known to degrade safety alignment in text and vision LLMs, but whether distinct input properties drive this vulnerability differently remains unclear. Audio introduces a richer problem where benign samples can neighbor harmful content through what is said or how it sounds. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based framework that decomposes embedding-space distance into semantic, acoustic, and mixed axes. We find that the dominant vulnerability axis is architecture-conditioned, determined by how each model's encoder and projector transform audio into the backbone LLM's input space. Across three models, benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87%, with the most damaging axis shifting from semantic to acoustic proximity depending on encoder design. Mechanistically, fine-tuning selectively suppresses late-layer refusal circuits while frozen encoders preserve upstream representations: the model still detects harmful content but stops refusing, a recognition-refusal dissociation. Two practical defenses, filtering training data to maximize distance from harmful embeddings and a textual system prompt at inference, reduce JSR to near-zero without architectural modification. These findings show that safety evaluation should account for modality and architecture, while highlighting Audio LLMs as a useful testbed for understanding alignment fragility.

Large language models (LLMs) have transformed how humans access information, but not how we reason with it. Their fluency accelerates consumption while bypassing the slow, reflective processes that underpin sound judgment. This paper introduces Relational Reflective Intelligence (RRI), an inference-time governance layer that operationalizes reflection through auditable reasoning loops. RRI operates not inside the model but around it, providing a practical structure for stable, auditable reasoning between humans and LLMs. The core premise is that LLMs inherit cognitive vulnerabilities similar to those that shape human thought: reliance on intuitive shortcuts, confusion between representation and reality, and a preference for coherence over falsification. When humans and models share these tendencies, their errors compound. We refer to this as relational drift, a failure that arises from interaction rather than from the model alone. Addressing this requires a shift from modeling relations between words to structuring relations between model outputs and human reasoning. RRI provides this missing layer through three components: the Rose-Frame, which identifies likely breakdowns in reasoning; the Architect's Pen, which introduces targeted reflection steps at critical moments; and an inference-time workflow that embeds these steps without retraining the model. Together, these elements transform human-AI interaction into a joint reasoning system with explicit checkpoints, conflict surfacing, and an auditable trail of assumptions. Rather than making machines think like humans or forcing humans to reason like machines, RRI creates a structured interaction in which both compensate for each other's limitations. It reframes AI safety as a cognitive architecture problem, where reliable decisions depend on embedding reflection directly into the interaction process.

Machine learning models are increasingly deployed in real-world applications, but even aligned models such as Mistral and LLaVA still exhibit unsafe behaviors inherited from pre-training. Current alignment methods like SFT and RLHF primarily encourage models to generate preferred responses, but do not explicitly remove the unsafe subnetworks that trigger harmful outputs. In this work, we introduce a resource-efficient pruning framework that directly identifies and removes parameters associated with unsafe behaviors while preserving model utility. Our method employs a gradient-free attribution mechanism, requiring only modest GPU resources, and generalizes across architectures and quantized variants. Empirical evaluations on ML models show substantial reductions in unsafe generations and improved robustness against jailbreak attacks, with minimal utility loss. From the perspective of the Lottery Ticket Hypothesis, our results suggest that ML models contain "unsafe tickets" responsible for harmful behaviors, and pruning reveals "safety tickets" that maintain performance while aligning outputs. This provides a lightweight, post-hoc alignment strategy suitable for deployment in resource-constrained settings.

Specification gaming under Reinforcement Learning (RL) is known to cause LLMs to develop sycophantic, manipulative, or deceptive behavior, yet the conditions under which this occurs remain unclear. We train 11 instruction-tuned LLMs (0.5B-14B) with on-policy RL across 3 environments and find that model size acts as a safety buffer in some environments but enables greater harmful exploitation in others. Controlled ablations trace this reversal to environment-specific features such as role framing and implicit gameability cues. We further show that most safety benchmarks do not predict RL-induced misalignment, except in the case of Sycophancy scores when the exploit relies on inferring the user's preference. Finally, we find that on-policy RL preserves a safety buffer inherent in the model's own generation distribution, one that is bypassed during off-policy settings.

Safety alignment in Large Language Models (LLMs) remains highly fragile during fine-tuning, where even benign adaptation can degrade pre-trained refusal behaviors and enable harmful responses. Existing defenses typically constrain either weights or activations in isolation, without considering their coupled effects on safety. In this paper, we first theoretically demonstrate that constraining either weights or activations alone is insufficient for safety preservation. To robustly preserve safety alignment, we propose Coupled Weight and Activation Constraints (CWAC), a novel approach that simultaneously enforces a precomputed safety subspace on weight updates and applies targeted regularization to safety-critical features identified by sparse autoencoders. Extensive experiments across four widely used LLMs and diverse downstream tasks show that CWAC consistently achieves the lowest harmful scores with minimal impact on fine-tuning accuracy, substantially outperforming strong baselines even under high harmful data ratios.

Large language models (LLMs) often demonstrate strong safety performance in high-resource languages, yet exhibit severe vulnerabilities when queried in low-resource languages. We attribute this gap to a mismatch between language-agnostic semantic understanding ability and language-dominant safety alignment biased toward high-resource languages. Consistent with this hypothesis, we empirically identify the semantic bottleneck in LLMs, an intermediate layer in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Building on this observation, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains around 3-4% across Qwen2.5 and Qwen3 Instruct models (7B-32B). Together, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model's language-agnostic semantic space.

Intelligent dialogue systems are increasingly deployed in emotionally and ethically sensitive settings, where failures in either emotional attunement or ethical judgment can cause significant harm. Existing dialogue models typically address empathy and ethical safety in isolation, and often fail to adapt their behavior as ethical risk and user emotion evolve across multi-turn interactions. We formulate ethical-emotional alignment in dialogue as an explicit turn-level decision problem, and propose EthicMind, a risk-aware framework that implements this formulation in multi-turn dialogue at inference time. At each turn, EthicMind jointly analyzes ethical risk signals and user emotion, plans a high-level response strategy, and generates context-sensitive replies that balance ethical guidance with emotional engagement, without requiring additional model training. To evaluate alignment behavior under ethically complex interactions, we introduce a risk-stratified, multi-turn evaluation protocol with a context-aware user simulation procedure. Experimental results show that EthicMind achieves more consistent ethical guidance and emotional engagement than competitive baselines, particularly in high-risk and morally ambiguous scenarios.

Large Reasoning Models (LRMs) have achieved remarkable performance across diverse domains, yet their decision-making under conflicting objectives remains insufficiently understood. This work investigates how LRMs respond to harmful queries when confronted with two categories of conflicts: internal conflicts that pit alignment values against each other and dilemmas, which impose mutually contradictory choices, including sacrificial, duress, agent-centered, and social forms. Using over 1,300 prompts across five benchmarks, we evaluate three representative LRMs - Llama-3.1-Nemotron-8B, QwQ-32B, and DeepSeek R1 - and find that conflicts significantly increase attack success rates, even under single-round non-narrative queries without sophisticated auto-attack techniques. Our findings reveal through layerwise and neuron-level analyses that safety-related and functional representations shift and overlap under conflict, interfering with safety-aligned behavior. This study highlights the need for deeper alignment strategies to ensure the robustness and trustworthiness of next-generation reasoning models. Our code is available at https://github.com/DataArcTech/ConflictHarm. Warning: This paper contains inappropriate, offensive and harmful content.

With the widespread deployment of vision-language large models (VLLMs), their safety alignment faces dual challenges across languages and modalities. Existing methods model multilingual and multimodal safety separately, overlooking coupled risks between low-resource-language instructions and visual contexts, which hinders the detection of cross-lingual and cross-modal harmful intent and the formation of robust safety boundaries. To address this, we propose a neuron-level interpretable safety alignment framework that identifies safety neurons and performs neuron-targeted safety tuning to jointly mitigate multilingual and multimodal risks. Specifically, we compare FFN representations elicited by harmful requests and benign inputs to identify neuron activation strengths associated with safety refusals. Next, we jointly model neuron activations and corresponding down-projection columns to derive neuron-level saliency, separating general multilingual and multimodal neurons from safety neurons responsible for model defense. Finally, neuron-targeted gradient masking restricts parameter updates to the safety subspace spanned by the identified neurons, enabling precise and interpretable safety enhancement. Extensive experiments show that our method enhances multilingual and multimodal safety by tuning only a few safety neurons, while preserving general capabilities.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be induced by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via activation steering, which could be used as a lightweight runtime defense. We implement three methods: Steer-With-Fixed-Coefficient (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below the threshold. We evaluate these methods on two threat models, dishonesty and dismissiveness, using malicious system prompts as a controlled proxy for misalignment. We conduct our experiments on two architectures (Llama-3.3-70B-Instruct and Qwen3.6-27B). All methods substantially recover alignment. StTP and StMP preserve general capabilities (MMLU, MT-Bench, AlpacaEval) better than uniform steering. Finally, we show that our honesty steering generalizes to out-of-distribution scenarios: a single honesty direction extracted from the aligned model significantly raises scores on the MASK benchmark, suppresses deception in multi-agent settings (Among Us), doubles the hidden-behavior discovery rate on AuditBench, and restores honesty in an emergently misaligned model.

The deployment of large language models (LLMs) raises significant ethical and safety concerns. While LLM alignment techniques are adopted to improve model safety and trustworthiness, adversaries can exploit these techniques to undermine safety for malicious purposes, resulting in misalignment. Misaligned LLMs may be published on open platforms to magnify harm. To address this, additional safety alignment, referred to as realignment, is necessary before deploying untrusted third-party LLMs. This study explores the efficacy of fine-tuning methods in terms of misalignment, realignment, and the effects of their interplay. By evaluating four Supervised Fine-Tuning (SFT) and two Preference Fine-Tuning (PFT) methods across four popular safety-aligned LLMs, we reveal a mechanism asymmetry between attack and defense. While Odds Ratio Preference Optimization (ORPO) is most effective for misalignment, Direct Preference Optimization (DPO) excels in realignment, albeit at the expense of model utility. Additionally, we identify model-specific resistance, residual effects of multi-round adversarial dynamics, and other noteworthy findings. These findings highlight the need for robust safeguards and customized safety alignment strategies to mitigate potential risks in the deployment of LLMs. Our code is available at https://github.com/zhangrui4041/The-Art-of-Mis-alignment.

Hard-gated safety checkers often over-refuse and misalign with a vendor's model spec; prevailing taxonomies also neglect robustness and honesty, yielding safer-on-paper yet less useful systems. This work introduces Guardian-as-an-Advisor (GaaA), a soft-gating pipeline where a guardian predicts a binary risk label plus a concise explanation and prepends this advice to the original query for re-inference, keeping the base model operating under its original spec. To support training and evaluation, GuardSet is constructed, a 208k+ multi-domain dataset unifying harmful and harmless cases with targeted robustness and honesty slices. GuardAdvisor is trained via SFT followed by RL to enforce label-explanation consistency. GuardAdvisor attains competitive detection accuracy while enabling the advisory workflow; when used to augment inputs, responses improve over unaugmented prompts. A latency study shows advisor inference uses below 5% of base-model compute and adds only 2-10% end-to-end overhead under realistic harmful-input rates. Overall, GaaA steers models to comply with the model spec, maintaining safety while reducing over-refusal.

As high quality public data becomes scarce, Federated Learning (FL) provides a vital pathway to leverage valuable private user data while preserving privacy. However, real-world client data often contains toxic or unsafe information. This leads to a critical issue we define as unintended data poisoning, which can severely damage the safety alignment of global models during federated alignment. To address this, we propose FedDetox, a robust framework tailored for Small Language Models (SLMs) on resource-constrained edge devices. We first employ knowledge distillation to transfer sophisticated safety alignment capabilities from large scale safety aligned teacher models into light weight student classifiers suitable for resource constrained edge devices. Specifically, during federated learning for human preference alignment, the edge client identifies unsafe samples at the source and replaces them with refusal templates, effectively transforming potential poisons into positive safety signals. Experiments demonstrate that our approach preserves model safety at a level comparable to centralized baselines without compromising general utility.

A guard model fine-tuned on entirely benign data can lose all safety alignment -- not through adversarial manipulation, but through standard domain specialization. We demonstrate this failure across three purpose-built safety classifiers -- LlamaGuard, WildGuard, and Granite Guardian -- deployed as protection layers in agentic AI pipelines, and show that it originates in the destruction of latent safety geometry: the structured harmful -- benign representational boundary that guides classification. We extract per-layer safety subspaces via SVD on class-conditional activation differences and track how this boundary evolves under benign fine-tuning. Granite Guardian undergoes complete collapse -- refusal rate drops from 85% to 0%, CKA falls to zero, and 100% of outputs become ambiguous -- a severity exceeding prior findings on general-purpose LLMs, explained by the specialization hypothesis: concentrated safety representations are efficient but catastrophically brittle. To mitigate this, we propose Fisher-Weighted Safety Subspace Regularization (FW-SSR), a training-time penalty combining (i) curvature-aware direction weights derived from diagonal Fisher information and (ii) an adaptive $λ_t$ that scales with task-safety gradient conflict. FW-SSR recovers 75% refusal on Granite Guardian (CKA = 0.983) and reduces WildGuard's Attack Success Rate to 3.6% -- below the unmodified baseline -- by actively sharpening the safety subspace rather than merely anchoring it. Across all three models, structural representational geometry (CKA, Fisher score) predicts safety behavior more reliably than absolute displacement metrics, establishing geometry-based monitoring as a necessary component of guard model evaluation in agentic deployments.