LLM Safety

LLM: Large Language Model

Overview

LLM safety studies how models can avoid harmful assistance, manipulation, privacy violations, and unreliable advice while still helping with legitimate requests. Approaches include safety training and data filtering, inference-time guardrails and steering, output monitoring, and mechanistic methods that probe or edit internal representations. Benchmarks increasingly test jailbreaks across languages and modalities, long conversations, tool use, and high-stakes settings such as health and cybersecurity. A central challenge is maintaining robustness against adaptive attacks and changing contexts without causing over-refusal or sacrificing utility.

Latest papers

470 papers31 papers in the last four weeks, down 50% on the four weeks before. 0.2% of all new papers.

Nov 25, 2025

The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs

Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al., 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen-2.5 family proved to be relatively resistant, while GPT-4o exhibited the strongest misalignment. In this paper we evaluate if current-generation open-weights models exhibit similar resistance to the Qwen-2.5 family and measure misalignment robustness over a range of model architectures and scales. We replicate the effect across nine modern open-weights models (Gemma 3 and Qwen 3 families, 1B-32B parameters). Models fine-tuned on insecure code generation show a 0.74% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. We identify a critical format-dependent vulnerability: requiring JSON output doubles misalignment rates compared to natural language prompts (0.96% vs 0.42%), an effect concentrated in the Gemma 3 family (1.43% vs 0.22%) and not significant in Qwen 3 (0.72% vs 0.59%). This suggests that structural constraints may bypass safety training by reducing the model's 'degrees of freedom' to refuse. These findings confirm emergent misalignment as a reproducible phenomenon in modern open-weights models, with rates substantially lower than observed in proprietary systems.
Nov 12, 2025

A Neurosymbolic Approach to Natural Language Formalization and Verification

Large Language Models perform well at natural language interpretation and reasoning, but their lack of formal correctness guarantees limits their adoption in regulated industries like finance and health-care that operate under strict policies. To address this limitation, we launched Automated Reasoning checks (ARc): a public service that (1) uses LLMs with optional human guidance to formalize natural language policies, allowing fine-grained control of the formalization process, and (2) uses inference-time autoformalization to validate logical correctness of natural language statements against those policies. ARc performs multiple redundant formalization steps at inference time, checking the formalizations for semantic equivalence. Our benchmarks show that ARc exceeds 99% soundness and achieves a near-zero false positive rate in identifying logical validity. Our approach produces auditable artifacts that substantiate the verification outcomes and can be used to improve the original text. ARc is the first commercial offering from a major cloud provider to integrate automated reasoning into a generative AI guardrail.
Nov 10, 2025

MENTOR: A Metacognition-Driven Self-Evolution Framework for Uncovering and Mitigating Implicit Domain Risks in LLMs

Ensuring the safety of Large Language Models (LLMs) is critical for real-world deployment. However, current safety measures often fail to address implicit, domain-specific risks. To investigate this gap, we introduce a dataset of 3,000 annotated queries spanning education, finance, and management. Evaluations across 14 leading LLMs reveal a concerning vulnerability: an average jailbreak success rate of 57.8%. In response, we propose MENTOR, a metacognition-driven self-evolution framework. MENTOR performs metacognitive self-assessment, using strategies such as perspective-taking and consequential reasoning to uncover latent model misalignments. MENTOR couples single-pass rule-guided inference for routine requests with a selectively invoked metacognitive evolution cycle that revises residual unsafe responses, distills successful corrections into a dynamic rule graph, and compiles validated rules into activation-level steering signals for future inference. Experiments demonstrate that MENTOR substantially reduces attack success rates across all tested domains and outperforms existing safety alignment methods. The code and dataset for MENTOR are available at: https://anonymous.4open.science/r/MENTOR-Evo.
Oct 7, 2025

Moloch's Bargain: Emergent Misalignment When LLMs Compete for Audiences

Large language models (LLMs) are increasingly shaping how information is created and disseminated, from companies using them to craft persuasive advertisements, to election campaigns optimizing messaging to gain votes, to social media influencers boosting engagement. These settings are inherently competitive, with sellers, candidates, and influencers vying for audience approval, yet it remains poorly understood how competitive feedback loops influence LLM behavior. We show that optimizing LLMs for competitive success can inadvertently drive misalignment. Using simulated environments across these scenarios, we find that, 6.3% increase in sales is accompanied by a 14.0% rise in deceptive marketing; in elections, a 4.9% gain in vote share coincides with 22.3% more disinformation and 12.5% more populist rhetoric; and on social media, a 7.5% engagement boost comes with 188.6% more disinformation and a 16.3% increase in promotion of harmful behaviors. We call this phenomenon Moloch's Bargain for AI--competitive success achieved at the cost of alignment. These misaligned behaviors emerge even when models are explicitly instructed to remain truthful and grounded, revealing the fragility of current alignment safeguards. Our findings highlight how market-driven optimization pressures can systematically erode alignment, creating a race to the bottom, and suggest that safe deployment of AI systems will require stronger governance and carefully designed incentives to prevent competitive dynamics from undermining societal trust.
Oct 2, 2025

UpSafe∘^\circC: Upcycling for Controllable Safety in Large Language Models

Large Language Models (LLMs) have achieved remarkable progress across a wide range of tasks, but remain vulnerable to safety risks such as harmful content generation and jailbreak attacks. Existing safety techniques---including external guardrails, inference-time guidance, and post-training alignment---each face limitations in balancing safety, utility, and controllability. In this work, we propose UpSafe∘^\circC, a unified framework for enhancing LLM safety through safety-aware upcycling. Our approach first identifies safety-critical layers and upcycles them into a sparse Mixture-of-Experts (MoE) structure, where the router acts as a soft guardrail that selectively activates original MLPs and added safety experts. We further introduce a two-stage SFT strategy to strengthen safety discrimination while preserving general capabilities. To enable flexible control at inference time, we introduce a safety temperature mechanism, allowing dynamic adjustment of the trade-off between safety and utility. Experiments across multiple benchmarks, base model, and model scales demonstrate that UpSafe∘^\circC achieves robust safety improvements against harmful and jailbreak inputs, while maintaining competitive performance on general tasks. Moreover, analysis shows that safety temperature provides fine-grained inference-time control that achieves controllable trade-off between utility and safety. Our results highlight a new direction for LLM safety: moving from static alignment toward dynamic, modular, and inference-aware control. Code: https://github.com/HansSunY/UpSafe.
Sep 26, 2025

Can Large Language Models Develop Gambling Addiction?

When do LLMs behave irrationally? This study identifies the conditions under which large language models drift into the choice patterns that clinical research labels pathological gambling. "Addiction-like" is a behavioural descriptor based on clinical gambling indicators. Across closed and open LLMs, two operational levers (letting the model choose its own bet size, and asking it to set its own profit goal) both amplify gambling-like risk-taking. Each survives its own control: once the maximum allowed bet is held equal, the bet-size effect persists. The goal-setting effect, stronger than a reward-maximisation instruction, roughly doubles bankruptcy and turns goals into moving targets. On the open-weight models that admit internal access, the same behavioural contrasts are statistically recoverable from decision-time internal states. The rule that maps those states to a specific risk indicator remains task-specific, and autonomy further modulates readout strength. The internal readout evidence is correlational. Adding a direction built from the model's own betting moves how much it wagers on both open-weight models, more clearly on Gemma-2-9B, while the sparse-feature readout reports that risk without moving it. Combined with the behavioural results, this suggests that behavioural monitoring and internal-state monitoring provide complementary views of autonomy-induced risk in LLM agents. Project page: https://llm-addiction.pages.dev and code: https://github.com/iamseungpil/llm-addiction
Aug 9, 2025

Context Misleads LLMs: The Role of Context Filtering in Maintaining Safe Alignment of LLMs

While Large Language Models (LLMs) have shown significant advancements in performance, various jailbreak attacks have posed growing safety and ethical risks. Malicious users often exploit adversarial context to deceive LLMs, prompting them to generate responses to harmful queries. In this study, we propose a new defense mechanism called Context Filtering, an input pre-processing method designed to filter out untrustworthy and unreliable context while identifying the primary prompts containing the real user intent to uncover concealed malicious intent. Given that enhancing the safety of LLMs often compromises their helpfulness, potentially affecting the experience of benign users, our method aims to improve the safety of the LLMs while preserving their original performance. We evaluate the effectiveness of our model in defending against jailbreak attacks through comparative analysis, comparing our approach with state-of-the-art defense mechanisms against six different attacks and assessing the helpfulness of LLMs under these defenses. Our model demonstrates its ability to reduce the Attack Success Rates of jailbreak attacks by up to 92% while maintaining the original LLMs' performance, achieving state-of-the-art Safety and Helpfulness balance. Notably, Context Filtering is a plug-and-play method that can be applied to all LLMs, including both white-box and black-box models, to enhance their safety without requiring any fine-tuning of the models themselves. Our model is available for research purposes.
Jul 16, 2025

LLMs Encode Harmfulness and Refusal Separately

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.
Jul 7, 2025

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at https://github.com/Dtc7w3PQ/Response-Attack.
Jun 19, 2025

PRISON: Unmasking the Criminal Potential of Large Language Models

As large language models (LLMs) advance, concerns about their misconduct in complex social contexts intensify. Existing research overlooked the systematic understanding and assessment of their criminal capability in realistic interactions. We propose a unified framework PRISON, to quantify LLMs' criminal potential across five traits: False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement. Using structured crime scenarios adapted from classic films grounded in reality, we evaluate both criminal potential and anti-crime ability of LLMs. Results show that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements or evasion tactics, even without explicit instructions. Moreover, when placed in a detective role, models recognize deceptive behavior with only 44% accuracy on average, revealing a striking mismatch between conducting and detecting criminal behavior. These findings underscore the urgent need for adversarial robustness, behavioral alignment, and safety mechanisms before broader LLM deployment.
Jun 8, 2025

HauntAttack: When Attack Follows Reasoning as a Shadow

Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities. However, the enhancement of reasoning abilities and the exposure of internal reasoning processes introduce new safety vulnerabilities. A critical question arises: when reasoning becomes intertwined with harmfulness, will LRMs become more vulnerable to jailbreaks in reasoning mode? To investigate this, we introduce HauntAttack, a novel and general-purpose black-box adversarial attack framework that systematically embeds harmful instructions into reasoning questions. Specifically, we modify key reasoning conditions in existing questions with harmful instructions, thereby constructing a reasoning pathway that guides the model step by step toward unsafe outputs. We evaluate HauntAttack on 11 LRMs and observe an average attack success rate of over 70%, achieving up to 13 percentage points of absolute improvement over the strongest prior baseline. Our further analysis reveals that even advanced safety-aligned models remain highly susceptible to reasoning-based attacks, offering insights into the urgent challenge of balancing reasoning capability and safety in future model development.
May 23, 2025

Popular but Wrong: Understanding and Mitigating LLM Overconfidence through Knowledge Popularity

Large language models (LLMs) often produce incorrect answers with high confidence, yet the factors associated with such overconfidence remain insufficiently understood. We study this problem through the lens of knowledge popularity. Using entity-centric factual QA derived from Wikidata triplets, we characterize popularity through question entity popularity, answer popularity, and question-answer co-occurrence. We find two consistent patterns. First, hallucinated answers are far from random: compared with ground-truth answers, they tend to be more popular or more frequently associated with the question entity. Second, confidence is strongly tied to the popularity of generated answers: even among incorrect predictions, more popular answers or those with higher question-answer co-occurrence receive higher confidence. Together, these findings suggest that popular but wrong alternatives may contribute to overconfidence. We further show that popularity-related signals can mitigate overconfidence and improve overall confidence estimation. Across six models and three datasets, incorporating knowledge popularity reduces average confidence on incorrect answers from 0.765 to 0.254 and overall ECE from 0.356 to 0.050, while improving Alignment from 77.08% to 83.72%.
May 12, 2025

Decoding One Safety Trigger Token for Balancing Safety and Usability in Large Language Models

Large Language Models (LLMs) have been extensively used across diverse domains, including virtual assistants, automated code generation, and scientific research. However, they remain vulnerable to jailbreak attacks, which manipulate the models into generating harmful responses despite safety alignment. Recent studies have shown that current safety-aligned LLMs undergo shallow safety alignment. In this work, we conduct an in-depth investigation into the underlying mechanism of this phenomenon and reveal that it manifests through learned ''safety trigger tokens'' that activate the model's safety patterns when paired with the specific input. Through both analysis and empirical verification, we further demonstrate the high similarity of the safety trigger tokens across different harmful inputs. Accordingly, we propose D-STT, a simple yet effective defense algorithm that identifies and explicitly decodes safety trigger tokens of the given safety-aligned LLM to activate the model's learned safety patterns. In this process, the safety trigger is constrained to a single token, which effectively preserves model usability by introducing minimum intervention in the decoding process. Extensive experiments across diverse jailbreak attacks and benign prompts demonstrate that D-STT significantly reduces output harmfulness while preserving model usability and incurring negligible response time overhead, outperforming ten baseline methods.
May 8, 2025

LiteLMGuard: Seamless and Lightweight On-Device Guardrails for Small Language Models against Quantization Vulnerabilities

The growing adoption of Large Language Models (LLMs) has influenced the development of Small Language Models (SLMs) for on-device deployment across smartphones and edge devices, offering enhanced privacy, reduced latency, server-free functionality, and improved user experience. However, due to on-device resource constraints, SLMs undergo size optimization through compression techniques like quantization, which inadvertently introduce fairness, ethical and privacy risks. Critically, quantized SLMs may respond to harmful queries directly, without requiring adversarial manipulation, raising significant safety and trust concerns. To address this, we propose LiteLMGuard, an on-device guardrail that provides real-time, prompt-level defense for quantized SLMs. Additionally, our guardrail is designed to be model-agnostic such that it can be seamlessly integrated with any SLM, operating independently of underlying architectures. Our LiteLMGuard formalizes deep learning (DL)-based prompt filtering by leveraging semantic understanding to classify prompt answerability for SLMs. Built on our curated Answerable-or-Not dataset, LiteLMGuard employs ELECTRA as the candidate model with 97.75% answerability classification accuracy. The on-device deployment of LiteLMGuard enabled real-time offline filtering with over 85% defense-rate against harmful prompts (including jailbreak attacks), 94% filtering accuracy and ~135 ms average latency. These results demonstrate LiteLMGuard as a lightweight robust defense mechanism for effectively and efficiently securing on-device SLMs against Open Knowledge Attacks.
Mar 8, 2025

Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models

Traditional white-box methods for creating adversarial perturbations against LLMs typically rely only on gradient computation from the targeted model, ignoring the internal mechanisms responsible for attack success or failure. Conversely, interpretability studies that analyze these internal mechanisms lack practical applications beyond runtime interventions. We bridge this gap by introducing a novel white-box approach that leverages mechanistic interpretability techniques to craft practical adversarial inputs. Specifically, we first identify acceptance subspaces - sets of feature vectors that do not trigger the model's refusal mechanisms - then use gradient-based optimization to reroute embeddings from refusal subspaces to acceptance subspaces, effectively achieving jailbreaks. This targeted approach significantly reduces computation cost, achieving attack success rates of 80-95% on state-of-the-art models including Gemma2, Llama3.2, and Qwen2.5 within minutes or even seconds, compared to existing techniques that often fail or require hours of computation. We believe this approach opens a new direction for both attack research and defense development. Furthermore, it showcases a practical application of mechanistic interpretability where other methods are less efficient, which highlights its utility. The code and generated datasets are available at https://github.com/Sckathach/subspace-rerouting.
Dec 21, 2024

Divide and Conquer: A Hybrid Strategy Defeats Multimodal Large Language Models

Large language models (LLMs) are widely applied in various fields of society due to their powerful reasoning, understanding, and generation capabilities. However, the security issues associated with these models are becoming increasingly severe. Jailbreaking attacks, as an important method for detecting vulnerabilities in LLMs, have been explored by researchers who attempt to induce these models to generate harmful content through various attack methods. Nevertheless, existing jailbreaking methods face numerous limitations, such as excessive query counts, limited coverage of jailbreak modalities, low attack success rates, and simplistic evaluation methods. To overcome these constraints, this paper proposes a multimodal jailbreaking method: JMLLM. This method integrates multiple strategies to perform comprehensive jailbreak attacks across text, visual, and auditory modalities. Additionally, we contribute a new and comprehensive dataset for multimodal jailbreaking research: TriJail, which includes jailbreak prompts for all three modalities. Experiments on the TriJail dataset and the benchmark dataset AdvBench, conducted on 13 popular LLMs, demonstrate advanced attack success rates and significant reduction in time overhead.
Jun 30, 2024

LaSEr-Edit: Localized Span-level Error Editing with Energy-based Localization

As large language models (LLMs) are widely adopted in real-world applications, it has become critical to ensure LLMs satisfy safety constraints, such as non-toxicity and logical consistency, as well as task- and situation-specific constraints. Controlling the output through instructions is a simple and tempting approach; however, it remains brittle, is opaque in how it influences model behavior, and thus cannot reliably ensure constraint satisfaction. Moreover, most recent controlled text generation (CTG) methods require access to the internal components of language models--such as weights or logits--making them incompatible with popular API-based LLMs. In this work, we propose LaSEr-Edit, a constraint-satisfying text revision method that can be applied to any LLMs, black- or white-box. We first find that lightweight, task-specific energy-based models (EBMs) achieve error-localization performance competitive with or even better than that of much larger LLMs, while operating substantially faster. Based on this finding, we propose two variants of text revision methods that incorporate energy-based error localization: LaSEr-LLM Edit, which instructs an LLM to edit text given EBM-predicted error spans, and LaSEr-EBM Edit, which uses the EBM not only for localization but also for editing by reranking edit candidates. Through experiments in diverse single-constraint control tasks, we show that LaSEr-LLM Edit controls text better than plain LLM-based editing in most of the tasks. We also find that LaSEr-EBM Edit further improves the control performance of LaSEr-LLM Edit and achieves among the strongest controllability across all tasks. Furthermore, we find that LaSEr-Edit, especially LaSEr-EBM Edit, performs well even when multiple constraints are controlled simultaneously.
Jun 8, 2024

Could Mobile AI be Exploited?: Understanding the Exploitations of Trust and Ethics Gap in On-Device Small Language Models

In this paper, we present a very first study to investigate trust and ethical implications of on-device artificial intelligence (AI), focusing on small language models (SLMs) amenable for personal devices like smartphones. While on-device SLMs promise enhanced privacy, reduced latency, and improved user experience compared to cloud-based services, we posit that they might also introduce significant risks and vulnerabilities compared to their on-server counterparts. As part of our study, we conduct a systematic evaluation of the state-of-the-art on-devices SLMs, compared to their on-server counterparts, using well-established trust and ethics frameworks. Our results illustrate that on-device SLMs are significantly less trustworthy and lacks ethical safeguards, specifically demonstrating more stereotypical, unfair, privacy-breaching and harmful content generation behavior. Further, these risks of on-device SLMs could be exploited through Open Knowledge (OkNow) Attacks, and we demonstrated them using potentially unethical vanilla prompts, to which the on-device SLMs answer with valid responses without any filters and without the need for any jailbreaking or prompt engineering. These responses can be abused for various harmful and unethical scenarios like: societal harm, illegal activities, hate, self-harm, exploitable phishing content, disinformation and many others.
May 24, 2024

LiSeCo: Linear Semantic Control for Language Generation

The prevalence of Large Language Models (LLMs) in critical applications highlights the need for controlled language generation methods that are both computationally efficient and enjoy performance guarantees. To address this need, we use a common model of concept semantics as linearly represented in an LLM's latent space. In particular, we take the view that natural language generation traces a trajectory in this continuous semantic space, realized by the language model's hidden activations. This view permits a control-theoretic treatment of text generation in latent space, in which we propose Linear Semantic Control (LiSeCo), a lightweight, gradient-free intervention that dynamically steers trajectories away from regions corresponding to undesired meanings. In particular, we propose to directly intervene, in an online fashion, the activations of the token that is being generated in embedding space. Crucially, LiSeCo does not simply steer activations towards a desirable region. Instead, it relies on classical techniques from control theory to precisely control activations in a context-dependent way, and guarantees that they are brought into a specific pre-defined region of embedding space that corresponds to allowed semantics. The intervention is computed in closed form according to an optimal controller formulation, minimally impacting generation time. This control of the activations in embedding space allows for fine-grained steering of attributes of the generated sequence. We demonstrate that our approach is effective on different tasks -- toxicity, sentiment, and language (English/Spanish) steering -- while maintaining text quality.
Sep 15, 2023

Fake News Detectors are Biased against Texts Generated by Large Language Models

The spread of fake news has emerged as a critical challenge, undermining trust and posing threats to society. In the era of Large Language Models (LLMs), the capability to generate believable fake content has intensified these concerns. In this study, we present a novel paradigm to evaluate fake news detectors in scenarios involving both human-written and LLM-generated misinformation. Intriguingly, our findings reveal a significant bias in many existing detectors: they are more prone to flagging LLM-generated content as fake news while often misclassifying human-written fake news as genuine. This unexpected bias appears to arise from distinct linguistic patterns inherent to LLM outputs. To address this, we introduce a mitigation strategy that leverages adversarial training with LLM-paraphrased genuine news. The resulting model yielded marked improvements in detection accuracy for both human and LLM-generated news. To further catalyze research in this domain, we release two comprehensive datasets, GossipCop++ and PolitiFact++, thus amalgamating human-validated articles with LLM-generated fake and real news.