Adversarial Robustness of Language Models

Adversaries can implant latent harmful behavior by poisoning as few as 1% of fine-tuning examples. The contamination is invisible to every output-level defense: harmful behavior lies dormant in the model's hidden-state geometry and does not appear in generated text until contamination exceeds 7.5%. We introduce CANARY (Contamination Auditor via Neural Activation Representation Yield), a zero-label checkpoint auditor that detects this hidden shift directly from two forward passes over an unlabeled prompt set. CANARY projects the hidden-state difference through a Sparse Autoencoder, filtering style noise to isolate meaningful semantic drift. It achieves AUROC = 1.000 at 1% contamination (95% CI = [0.997, 1.000]; Cohen's d = 3.28) across four model architectures and two training paradigms, 7.5x below where any output-level method fires, with zero false positives on benign fine-tuning and full robustness to style-matching and gradient-noise adaptive attacks. The same SAE feature basis drives a complete governance pipeline: SAE-filtered amplification surfaces latent harm at a 5x higher rate than standard generation; score-ranked prompts yield 4.2x red-teaming lift; and suppressing a handful of contamination-specific features at inference time reduces harm from 70% to 10% with no perplexity penalty. CANARY is the first zero-label framework to detect, verify, prioritize, and remediate supply-chain contamination from hidden states alone.

Large language model (LLM) agents increasingly rely on reusable skills i.e. documents describing task-specific procedures. However, this introduces a new attack surface for agents to manage. We study two complementary directions for this threat. First, we evaluate guardian-based defenses: an intermediary LLM agent that acts as a mediator for skill file access (dynamic guardian) or pre-rewrites these files at build time (static guardian). Across three LLM agent families, our guardians cut attack success rate (ASR) by well over half while preserving task utility. Second, we stress test them through attack reframing using four attacks that preserve the malicious instruction but change the phrasing. For non-guardian setup, the reframing pushes the ASR up to 81.4%, but the dynamic guardian brings it down to 18.6%, showing that real-time mediation is a robust defense.

Low-rank adaptation has become a standard method for parameter-efficient fine-tuning of large language models, but even small amounts of unsafe or adversarial fine-tuning data can substantially weaken the safety behavior of aligned models. Existing safety-preserving LoRA methods often rely on hard interventions such as projection, pruning, thresholding, or additional training objectives. While these methods can suppress unsafe update directions, they may also remove task-relevant information or require extra tuning. We introduce CSULoRA, a post-hoc method for correcting trained LoRA adapters through closest safe update estimation. CSULoRA estimates a safety-aligned subspace from the weight displacement between a safety-aligned model and its corresponding base checkpoint. It then decomposes each LoRA update into fully aligned, partially aligned, and off-subspace components. Instead of discarding components outside the estimated safety subspace, CSULoRA solves a closed-form penalized minimum-change problem that preserves the fully aligned component while smoothly attenuating potentially unsafe directions according to their relative energy. In adversarial fine-tuning experiments, CSULoRA substantially reduces attack success rate while preserving most of the utility gains obtained from standard LoRA fine-tuning.

Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers. Under sustained adversarial pressure we find a previously undocumented failure mode: the chain-of-thought stays factually correct from first turn to last while the emitted answer flips wrong. We call this unfaithful capitulation (UC) and isolate it with a $2\times 2$ latent-versus-behavioral framework that flip-rate metrics and single-turn faithfulness probes both miss. Across three datasets (MT-Consistency, MMLU-Pro, GSM8K), the latent-correct rate at the behavioral flip clusters near 50% in think mode and collapses to 11-15% under no_think -- paired, within-model causal evidence that reasoning creates the gap. Across models the effect tracks the reasoning channel (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it). An independent GPT-4o judge corroborates $86\%$ of UC labels; a token-level probe shows the answer-slot argmax is correct in $84\%$ of UC cells; and a naive trace-anchored defense backfires. We release all trajectories, traces, and judge labels.

As LLMs gain stronger reasoning capabilities, their extended chain-of-thought introduces new degrees of complexity for defending against adversarial jailbreaks and prompt injection. We study consistency training, a family of fine-tuning objectives that enforce identical behavior on clean prompts and adversarial rewrites, and evaluate its two main variants, output-level (BCT) and activation-level (ACT), across five reasoning models. We formulate both methods as a prompt injection defense and find ACT to be competitive with other training-based defenses while requiring only self-supervised pairs of clean and wrapped prompts. Our experiments also generalize both techniques within the jailbreak setting, demonstrating that ACT remains more robust to adaptive attacks. We also provide mechanistic evidence that ACT's defense against jailbreaks is encoded as a roughly linear shift in activation space at the assistant-turn boundary. After ACT training, we can recover a single steering direction that controls refusal on reasoning models with minimal effect on benign inputs. We find that ACT remains robust even when the model's chain-of-thought is replaced with a compliant trace from the undefended base model, pivoting to refuse prefilled jailbreaks. Together, these results suggest that supervising internal representations is a surprisingly effective and interpretable approach to various forms of safety training in reasoning models.

Large language models(LLMs) increasingly match expert performance on licensing examinations, yet routine clinical use remains limited because governance requires auditable reasoning, safety and ethics alignment, and resilience to adversarial misuse. Here we present SafeMed-R1, trained with a traceable Clinical Trust Signals(CTS) pipeline that links each reasoning instance to clinician rubric scores and edit histories, and aligned through safety and ethics supervision and red team stress testing. SafeMed-R1 attains a macro-averaged accuracy of 79.6% across clinical benchmarks. Under adversarial safety testing, it shows the lowest aggregated risk and reduces unsafe outputs by about 3 to 5% relative to its baseline. In a paired expert study of 30 medication safety vignettes, SafeMed-R1 matches PGY1 and PGY2 residents on medical correctness and scores higher for medication safety, guideline consistency, and clinical usefulness. Collectively, these results suggest that clinician-audited supervision provenance, together with domain-tailored safety and ethics alignment, can strengthen governance-relevant evidence without relying on inference-time retrieval or citation grounding.

Large Language Models (LLMs) are increasingly vulnerable to adversarial prompts that exploit semantic ambiguities to bypass safety mechanisms, resulting in harmful or inappropriate outputs. Such attacks, including jailbreaking and prompt injection, pose significant risks to the integrity and availability of LLMs in security-critical applications. This paper proposes the Adversarial Prompt Disentanglement (APD) framework, a novel defense mechanism that proactively identifies and neutralizes malicious components in input prompts before they are processed by the LLM. The APD framework integrates three key innovations: (1) a mutual information-based semantic decomposition method to isolate adversarial and benign prompt components, ensuring statistical independence; (2) a graph-based intent classification approach that leverages spectral analysis to detect malicious patterns in prompt semantics; and (3) a lightweight transformer-based classifier trained on real-world datasets of toxic and jailbreaking prompts, enabling efficient and accurate adversarial intent detection. Evaluated on diverse datasets containing adversarial prompts, APD demonstrates superior robustness, reducing harmful output generation by over 85% while maintaining negligible impact on model performance. The framework's computational efficiency supports real-time deployment, making it a practical solution for securing LLMs. Our work addresses critical challenges in machine learning security on novel attacks and integrity methods for ML systems, and offers a scalable, ethically grounded defense against prompt-based adversarial threats.

Legal reasoning requires distinguishing changes that matter from those that do not. Legal AI should remain stable under legally irrelevant perturbations, but should change when perturbations alter legally material points. We formulate this requirement as a legal-relevance-sensitive evaluation problem: LLMs should only be sensitive to the legally relevant change. We introduce a unified evaluation suite covering should-change and should-not-change evaluation across judicial fairness, robustness, and statute-confusion scenarios. Our evaluation shows that existing legal LLMs are systematically sensitive to legally irrelevant variations and often fail to distinguish related legal elements and statutory rules. To mitigate these failures, we present LexGuard, an adversarial multi-agent framework grounded in formal reasoning. LexGuard formalizes statutes into executable constraints, uses adversarial agents to extract competing fact-statute arguments, and invokes SMT solvers to verify legal satisfaction and logical consistency. Experiments show that LexGuard improves legal reasoning reliability by reducing vulnerability to manipulative framing, improving disambiguation among similar statutes, limiting the influence of legally irrelevant attributes, and increasing consistency under benign reformulations. We show that legal trustworthiness requires not only accuracy, but calibrated sensitivity to legally material changes.

Direct Preference Optimisation (DPO) is widely used for safety alignment in large language models. However, prior work shows it is brittle and exhibits poor out-of-distribution (OOD) generalisation. In this paper, we investigate whether Curriculum Learning can improve the robustness of DPO-based safety alignment. We propose Staged-Competence, a curriculum-based framework that organises preference data by difficulty, employs competence-based sampling, and progressively updates the reference model during training. Averaged across three model families, Staged-Competence reduces OOD harmful response rates by 16% and jailbreak attack success rates by 20%, while preserving general capabilities with near-zero over-refusal. We further show that Staged-Competence (1) matches baseline safety with only 75% of the training data and (2) yields better separation between safe and unsafe responses. Staged-Competence is agnostic to the policy optimisation loss and can extend to other DPO variants and alignment domains. Our code and data are available at https://github.com/Sandeep5500/curriculum-learning-for-safety.

Large language models are widely deployed in high-stakes NLP tasks, yet risks such as bias, hallucination, adversarial vulnerability and unreliable generalization remain. Probe-based auditing reveals inconsistencies in model behavior. Adversarial text generation uncovers robustness gaps, especially in lower-resourced languages with limited benchmarks. Enterprise text-to-SQL settings expose the difficulty of validating outputs over private and large-scale databases. Human supervision is essential for probe validation, adversarial verification and domain-specific annotation, but it is costly and hard to scale. This survey examines recent human-in-the-loop methods that shift NLP from automation toward collaboration for safety and trustworthiness. We review how human expertise supports auditing, robustness evaluation, data construction and model steering. Our findings highlight gaps in scalable probing, sustainable robustness benchmarks, low-resource settings and governance of private systems. We outline practical research directions for adaptive auditing, collaborative evaluation and accountable deployment.

Large language model (LLM) safety classifiers such as Llama Guard are effective at detecting overtly harmful prompts but remain vulnerable to adversarial jailbreak attacks that disguise malicious intent through role-play scenarios, fictional framing, and indirect requests. We present Reflect-Guard, a method that augments LLM-based safety classifiers with chain-of-thought self-reflection capabilities through parameter-efficient fine-tuning. Our approach distills analytical reasoning from GPT-4o-mini into structured reflection annotations, then trains Llama-Guard-3-8B via QLoRA to generate logical self-reflections before issuing safety verdicts. Using only 1000 training examples and updating just 0.5% of model parameters (~42M), Reflect-Guard achieves substantial improvements on two challenging benchmarks. On WildGuardTest, F1 score improves from 0.770 to 0.842 (+7.2 pp), with recall on adversarial prompts increasing from 0.513 to 0.921 (+40.8 pp). On JailbreakBench, the attack success rate drops from 10.3% to 1.8%, representing an 82.5% relative reduction. These gains are especially pronounced on adversarial inputs, where the explicit reasoning step enables the model to see through obfuscation techniques that defeat standard pattern-matching approaches. Our results demonstrate that teaching safety classifiers to reason about adversarial intent, rather than simply classify surface patterns, is a promising direction for robust LLM safety.

Jailbreak prompts can trigger harmful completions on aligned LLMs, In accordance, safety steering has been proposed: test-time activation interventions that steer jailbreak activations to trigger refusal while preserving benign utility. However, existing steering methods are fundamentally supervised and tied to a static, limited training set, whereas real jailbreaks evolve and are often out-of-distributed from the training set, leading to failures on unseen attacks. In this paper, we tackle the failure on unseen jailbreaks problem, base on unsupervised latent direction discovery. We propose a bi-level adversarial training framework for zero-shot jailbreak defense. In the inner step, we simulate diverse jail-broken activations by extrapolating from refusal-state harmful-request activations via unsupervised latent direction discovery, which expands the coverage of real jailbreak activation subspaces. In the outer step, we train a potential-induced steering field to push these adversarial jailbroken states into refusal regions while keeping benign unchanged. Across three LLMs and six classical jailbreak families, our method achieves strong defense with attack success rates mostly below 5%, and rising subspace coverage throughout training helps explain the improved generalization.

Large language models (LLMs) require robust toxicity evaluation beyond explicit wording. This setting remains underexplored in Chinese, where toxicity may combine semantic indirectness with surface obfuscation. We introduce Chinese Implicit Toxicity Attack (CITA), a controlled red-team evaluation and defense-data generation framework, not a deployable evasion tool. CITA uses three stages: (i) Harmful Intent Learning, (ii) Implicit Toxicity Enhancement, and (iii) Obfuscation Variant Rewriting, to preserve harmful intent, increase implicitness, and add controlled surface variants. On CITA-generated evaluation samples, the seven tested detectors exhibit substantial missed-detection risks, reaching an average ASR of 69.48%; human evaluation further confirms preserved harmfulness and increased implicitness/evasiveness. As a downstream defense application, we fine-tune a Chinese Implicit Toxicity Defense model (CITD) with CITA-generated red-team data, showing that such data can improve robustness through additional training.

Aligned models can misbehave in several ways: they are often sycophantic, fall victim to jailbreaks, or fail to include appropriate safety warnings. Consistency training is a promising new alignment paradigm to mitigate such failures by training invariants into the model using contrastive input pairs. Existing consistency training procedures generate the supervision signal once, offline, and use supervised fine-tuning (SFT) to update the model. Unfortunately, the resulting models tend to merely memorize the surface forms of the training distribution and thus generalize poorly and regress in their capabilities. We introduce On-Policy Consistency Training (OPCT), a new consistency training approach where the objective is computed over the model's own responses to prompts, supervised by itself conditioned on corresponding contrastive prompts. We evaluate OPCT on three safety axes: sycophancy, jailbreaking, and safety awareness. Across three model families, OPCT outperforms its SFT counterpart on all safety desiderata. It nearly halves the sycophancy rate relative to baseline (8.1% vs. 15.4%, compared to 11.2% for SFT). Under an adaptive per-target attacker, OPCT holds jailbreak defense success near 99% on held-out jailbreak behaviors, whereas SFT achieves 87% on average. On safety awareness, OPCT outperforms SFT in two out of three models, and matches it on the other. OPCT also largely avoids the capability regressions that SFT induces, such as a 28-point drop on MATH-500. Our results suggest that consistency training is best implemented as OPCT rather than as SFT, especially when generalization beyond the training distribution is desired.

Preference-based post-training aligns LLMs with human intent, yet safety behavior often remains brittle. A model may refuse a harmful request in a standard prompt but comply when the same intent is wrapped in adversarial wording. We suggest that robust safety requires context-invariant alignment, where behavior depends on the underlying intent rather than surface form. Enforcing invariance is difficult in alignment because not all training signals are equally trustworthy; for some prompt variants we can obtain verifiable feedback (e.g., multiple-choice), while for open-ended variants we typically rely on noisy, gameable reward proxies (e.g., learned judges). As a result, standard symmetric invariance regularizers can reduce cross-context discrepancies by lowering performance on reliable variants instead of improving open-ended robustness. To address this, we introduce Anchor Invariance Regularization (AIR), which treats verifiable prompts as anchors and uses a stop-gradient target to regularize only the open-ended variants toward the anchor performance. AIR is implemented as a plug-in auxiliary loss and combined with group-based preference optimization (e.g., GRPO) via heterogeneous prompt grouping. Across Safety, Moral Reasoning, and Math, AIR improves context invariance, boosting in-distribution group accuracy by 12.71% and out-of-distribution consistency by 33.49%, making safety constraints robust to adversarial framings.

While Large Language Models (LLMs) demonstrate remarkable capabilities, they remain susceptible to sophisticated, multi-step jailbreak attacks that circumvent conventional surface-level safety alignment by exploiting the internal generation process. To address these vulnerabilities, we propose Reflector, a principled two-stage framework that internalizes self-reflection within the generation trajectory. Reflector first leverages teacher-guided generation to produce high-quality reflection data for supervised fine-tuning (SFT), establishing structured reflection patterns. It subsequently uses Reinforcement Learning (RL) with outcome-driven and reward-validity supervision to instill robust, autonomous self-reflection capabilities. Empirical results show that Reflector achieves Defense Success Rates (DSR) exceeding 90% against complex indirect attacks while generalizing robustly across diverse threat scenarios. Notably, the framework enhances both task-specific and general utility, yielding a 5.85% gain on GSM8K alongside improved performance on knowledge-intensive benchmarks. By internalizing trajectory-level safety, Reflector overcomes the fundamental limitations of surface alignment without significant computational overhead, offering an efficient and scalable solution for the development of safe and capable LLMs.

Recent work has demonstrated the potential of contrastive steering for jailbreaking Large Language Models (LLMs). However, existing methods rely on limited and inherently biased contrastive prompts and require laborious manual tuning of steering strength, limiting their robustness and effectiveness. In this paper, we leverage the idea of model extraction to guide the learned steering vectors to approximate the ideal one and propose tuning the steering strength adaptively based on contrastive activations' statistics. Experiments demonstrate that our method notably improves the effectiveness and robustness of probe-based steering, without any extra contrastive prompts or laborious manual tuning. Being an attack paper, this paper focuses on revealing the breakdown of fortified LLMs, raising the average harmfulness score from 6% to 70%. Our code is available at https://github.com/fhdnskfbeuv/adaptiveSteering.

As jailbreaks, adversarially crafted inputs that bypass safety constraints, continue to be discovered in Large Language Models, practitioners increasingly rely on fine-tuning as a defensive strategy. Yet the theoretical foundations underlying this robustness fine-tuning remain underexplored. We introduce a game-theoretic framework in which the interaction between an evaluator (auditing the model for jailbreaks) and a trainer is formalized as a two-player game. A key feature of our approach is the use of group actions, a mathematical structure that captures symmetries and transformations, to formally represent data augmentation. The simplest non-trivial instance is the circle with cyclic translation groups, where we exhibit various regimes depending on the trainer's generalization range. Below a critical threshold, the evaluator maintains a constant miss ratio for linearly many rounds, whereas other settings can yield very different behaviors. We further provide empirical evidence supporting locality-dependence of the model: for the three model families we tested (Llama, Qwen and Mistral), we have significant evidence that fine-tuning on adversarial prompts induces only local generalization, with refusal rates on test examples highly correlated with the distance to the fine-tuning prompts. Our framework recasts the central object of adversarial evaluation: a benchmark is not a static set of prompts but an orbit under the evaluator's group action, and audit protocols that ignore trainer-side adaptation cannot distinguish a genuine fix from a memorized patch.

Model providers increasingly release the weights of large language models. Although these models are safety-aligned before release, their safeguards can often be removed by fine-tuning on harmful data. A growing class of defenses aims to make alignment robust to such malicious fine-tuning, but these defenses are typically evaluated against attacks with a fixed training budget, even though an attacker who holds the weights can simply train for longer. We ask whether current defenses withstand this simplest escalation. Surveying fifteen recent defenses, we find that they share a common weakness: each is built around a limited model of the attacker, such as a bounded perturbation, a short simulated attack, or a trained link between harmful and benign behavior, and nothing enforces that protection once the weights are released. We then test six representative defenses on four open-weight models by continuing the same harmful-only fine-tuning attack for three epochs and measuring harmfulness and capability along the way. In all 72 defended runs, the model is more harmful at the end of training than at release, and on Llama-3.1 at the highest learning rate the defended models end almost as harmful as the undefended one. The defenses are not equally weak: one defense kept harmfulness low on one model, and some attacks recovered harmfulness only at the cost of general capability. Current defenses can delay or disrupt malicious fine-tuning, but in most cases their measured resistance does not persist under continued training, and they should not yet be treated as durable protection.

Safety alignment often improves robustness to harmful queries at the cost of reasoning ability, a tradeoff known as the safety tax. A common cause is distributional mismatch: supervised fine-tuning trains the target model on safety demonstrations produced by humans, external models, or fixed self-generated traces, rather than on trajectories sampled from its own policy. We identify off-policy training mismatch as a second source of this tax and study on-policy self-distillation for safety alignment, which we call OPSA. The model generates its own rollouts and receives dense per-token KL supervision from a frozen teacher copy of itself conditioned on a privileged safety context. Because this teacher must be safer than the sampled student trajectory, we introduce \emph{teacher flip rate}: a criterion that measures how often a privileged context converts unsafe responses into safe ones. We use this signal to search for contexts that activate latent safety reasoning rather than merely elicit safe-looking demonstrations. Across two reasoning-model families and five model scales, OPSA achieves a stronger safety--reasoning tradeoff than off-policy self-distillation and external-teacher distillation under matched data and full-parameter fine-tuning, with the largest gains on smaller models (+8.85 points on R1-Distill-1.5B and +5.49 points on Qwen3-0.6B). The gains persist across training-set sizes and adaptive jailbreak evaluations. Token-level analyses further show that OPSA concentrates updates near early compliance-decision tokens, providing a mechanism for improving safety while preserving general reasoning.

Large Language Models (LLMs) trained on web-scale corpora inherently absorb toxic patterns from their training data. This leads to toxic degeneration where even innocuous prompts can trigger harmful outputs. This phenomenon poses significant risks for real-world deployments. Thus, necessitating effective mitigation strategies that should maintain model utility while ensuring safety. In this comprehensive replication study, we evaluate the efficacy of DExperts (Decoding-time Experts), which is an inference-time mitigation technique that steers generation without requiring model retraining. We structured our research into three systematic phases: (1) establishing baseline toxicity measurements using RealToxicityPrompts on standard GPT-2 models; then (2) implementing and evaluating DExperts to mitigate explicit toxicity; and finally (3) stress-testing the method against implicit hate speech using the adversarial ToxiGen dataset. Our empirical results confirm that while DExperts achieves near-perfect safety rates (100%) on explicit toxicity benchmarks, it exhibits brittleness against adversarial, implicit hate speech, with safety rates dropping to 98.5%. Furthermore, we quantify a critical trade-off. The method introduces a 10x latency penalty (from 0.2s to 2.0s per generation), posing challenges for real-time deployment scenarios. This study contributes to the growing body of work on AI safety by highlighting the robustness gap between explicit and implicit toxicity mitigation. We emphasize the need for more sophisticated approaches that generalize across diverse hate-speech patterns without incurring prohibitive computational costs.

Automated red-teaming for LLMs often discovers narrow attack slices, missing diverse real-world threats, and yielding insufficient data for safety fine-tuning. We introduce Persona-Conditioned Adversarial Prompting (PCAP), which conditions adversarial search on diverse attacker personas (e.g., doctors, students, malicious actors) and strategy sets to explore realistic attack scenarios. By running parallel persona-conditioned searches, PCAP discovers transferable jailbreaks across different contexts and generates rich defense datasets with automatic metadata tracking. On GPT-OSS 120B, PCAP increases attack success from 57% to 97% while producing 2-6$\times$ more diverse prompts covering varied real-world scenarios. Critically, fine-tuning lightweight adapters on PCAP-generated data significantly improves model robustness (recall: 0.36 $\rightarrow$ 0.99, F1: 0.53 $\rightarrow$ 0.96) with minimal false positives, demonstrating a practical closed-loop approach from vulnerability discovery to automated alignment.

We present Seirênes, a self-play RL framework that transforms contextual interference from a failure mode of LLM reasoning into an internal training signal for co-evolving more resilient reasoners. While RL with verifiable rewards has significantly advanced reasoning capabilities, models can still exhibit fragility when encountering non-idealized contexts: scenarios characterized by superfluous information, tangential instructions, or incidental correlations that differ from the clean distributions typical of standard benchmarks. Seirênes harnesses this vulnerability through a parameter-shared and adversarial self-play loop. Within this framework, a single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots, and solve problems by discerning the essential task from these perturbations to recover the core underlying logic. By pitting these competing objectives against each other, Seirênes compels the model to move beyond superficial pattern matching and anchors its capabilities in robust underlying reasoning. This continuous interaction sustains an informative co-evolutionary curriculum as the model improves. Across seven mathematical reasoning benchmarks and model scales from 4B to 30B, Seirênes achieves average gains of +10.2, +9.1, and +7.2 points. Besides, distracting contexts produced by the 4B Seirênes model reduce the accuracy of top-tier closed-source models (GPT and Gemini) by roughly 4--5 points, revealing Seirênes' general ability to uncover reasoning models' blind spots.

Large Reasoning Models possess remarkable capabilities for self-correction in general domain; however, they frequently struggle to recover from unsafe reasoning trajectories under adversarial attacks. Existing alignment methods attempt to mitigate this vulnerability by fine-tuning the model on expert data including reflection traces or adversarial prefixes. Crucially, these approaches are often hindered by static training data which inevitably deviate from model's dynamic, on-policy reasoning traces, resulting in model hardly covering its vast generation space and learning to recover from its own failures. To bridge this gap, we propose Self-ReSET, a pure reinforcement learning framework designed to equip LRMs with the intrinsic capacity to recover from their own safety error trajectories, which are subsequently reused as an initial state for reinforcement learning. Extensive experiments across various LRMs and benchmarks demonstrate that Self-ReSET significantly enhances robustness against adversarial attacks especially out-of-distribution (OOD) jailbreak prompts while maintaining general utility, along with efficient data utilization. Further analysis reveals that our method effectively fosters self-recovery patterns, enabling models to better identify and recover from unsafe intermediate error states back to benign paths. Our codes and data are available at https://github.com/Ing1024/Self-ReSET.

While explicit Chain-of-Thought (CoT) empowers large reasoning models (LRMs), it enables the generation of riskier final answers. Current alignment paradigms primarily rely on externally enforced compliance, optimizing models to detect malicious prompts rather than evaluating the safety of their own outputs. We argue that this approach remains largely behavioral: our empirical analysis reveals that ostensibly aligned models lack intrinsic safety understanding, often failing to verify their own response safety and remaining vulnerable to adversarial jailbreaks. To address this fundamental limitation, we propose Safety Internal (SInternal), a framework that internalizes safety specifications by training LRMs exclusively on safety verification tasks to critique their own generated answers using expert reasoning trajectories. We demonstrate that learning to verify induces a strong generalization for response safety, significantly enhancing robustness against out-of-domain jailbreaks. Furthermore, when combined with reinforcement learning, SInternal serves as a superior initialization compared to standard supervised fine-tuning, suggesting that internalizing safety understanding creates a more robust foundation for alignment than merely mimicking safe behaviors. Our codes are available at https://github.com/AlphaLab-USTC/SInternal

Current adversarial robustness methods for large language models require extensive datasets of harmful prompts (thousands to hundreds of thousands of examples), yet remain vulnerable to novel attack vectors and distributional shifts. We propose Latent Personality Alignment (LPA), a sample-efficient defense that achieves robustness by training models on abstract personality traits rather than specific harmful behaviors. Using fewer than 100 trait statements and latent adversarial training, LPA achieves comparable attack success rates to methods trained on 150k+ examples, while maintaining superior utility. Critically, LPA generalizes better to unseen attack distributions, reducing misclassification rates by 2.6x compared to baseline across six harm benchmarks -- without ever seeing harmful examples during training. Our results demonstrate that personality-based alignment offers a principled approach to building robust defenses with minimal cost.

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

Reinforcement learning from verifiable emotion rewards RLVER has produced language models with strong empathetic performance, evaluated on benchmarks that assume cooperative, honest users. Yet real emotional interactions systematically violate this assumption: users gaslight, escalate, and pressure AI systems for unconditional validation, dynamics that cooperative benchmarks cannot surface. We construct the Adversarial Empathy Benchmark AEB and introduce the Emotional Consistency Score ECS to evaluate empathetic robustness under adversarial conditions. AEB comprises six psychologically grounded adversarial trajectory types with discriminative reward structures that penalize formulaic responses; ECS formally disentangles a model's capacity to track user emotional states from its capacity to improve them. In a controlled experiment across eight scenario-matched conditions (think and no-think conditions on 2 RLVER models, and 2 base models (Qwen 1.5B and 7B) with 480 adversarial dialogues), RLVER-PPO-Think substantially outperforms the same-scale untuned baseline (0.963 vs. 0.761, $p<0.001, r=0.688$), with zero dialogue collapses and 47% higher hidden-intention detection. However, ECS remains nearly flat and is not significantly different for RLVER-PPO-Think versus Base-7B-Think ($p=0.650$): RL training improves emotional responsiveness without measurable gains in observable state tracking. We interpret the ECS--FS (Final Score) gap as a behavioral/legibility dissociation inside this simulator family, not as evidence about internal understanding or clinical readiness.

Motivated by the challenge to improve the adversarial robustness, security, and trust of medical decision making intelligent agents, this study develops a full-link security enhancement framework, which describes "input risk perception - medical evidence constraint - knowledge consistency verification - decision confidence reweighting - security output control - adversarial feedback update." We propose ARSM-Agent and define a weighted joint objective consisting of decision accuracy loss, adversarial robustness loss, safety refusal loss, and knowledge consistency loss, with weights of 0.3, 0.3, 0.2, and 0.2, respectively. The whole medical decision formulation is implemented by multi-module collaborative linkage. We verify that the algorithm is more efficient than four baselines, including LLM-Agent, Retrieval-Agent, Filter-Agent, and Adv-Train-Agent. Under semantic perturbation, prompt injection, drug-name confusion, and false-evidence attacks, ARSM-Agent reduces the overall attack success rate to 8.7% and achieves a knowledge consistency score of 0.91. Ablation experiments quantify each module's contribution: removing risk perception, evidence retrieval, consistency verification, and confidence reweighting reduces accuracy by 6.7%, 9.1%, 7.6%, and 4.4%, respectively, and increases attack success rate by 13.8%, 11.1%, 8.6%, and 6.9%. The proposed approach addresses key security issues of medical decision making intelligent agents, obtains secure decision making in challenging scenarios, and provides reliable intelligent support for medical decision-making intelligent agents.

The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.