This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs). By assuming knowledge of the trigger, we identify a specific ``trigger direction'' in the model's activations that corresponds to the internal representation of the trigger. We confirm the causal role of this linear direction by showing that interventions in both activation and parameter space consistently modulate the model's backdoor behavior across multiple datasets and attack types. Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers. Our analysis reveals distinct qualitative differences: static-patch triggers follow a different internal logic than stealthy, distributed triggers. We further examine the link between backdoors and adversarial attacks, specifically testing whether PGD-based perturbations (de-)activate the identified trigger mechanism. Finally, we propose a data-free, weight-based detection scheme for stealthy-trigger attacks. Our findings show that mechanistic interpretability offers a robust framework for diagnosing and addressing security vulnerabilities in computer vision.

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.

Membership inference attacks (MIAs) threaten the privacy of machine learning models by revealing whether a specific data point was used during training. Existing MIAs often rely on impractical assumptions such as access to public datasets, shadow models, confidence scores, or training data distribution knowledge and making them vulnerable to defenses like confidence masking and adversarial regularization. Label-only MIAs, even under strict constraints suffer from high query requirements per sample. We propose a cost-effective label-only MIA framework based on transferability and model extraction. By querying the target model M using active sampling, perturbation-based selection, and synthetic data, we extract a functionally similar surrogate S on which membership inference is performed. This shifts query overhead to a one-time extraction phase, eliminating repeated queries to M . Operating under strict black-box constraints, our method matches the performance of state-of-the-art label-only MIAs while significantly reducing query costs. On benchmarks including Purchase, Location, and Texas Hospital, we show that a query budget equivalent to testing $\approx1\%$ of training samples suffices to extract S and achieve membership inference accuracy within $\pm1\%$ of M . We also evaluate the effectiveness of standard defenses proposed for label-only MIAs against our attack.

With the rapid advancement and adoption of Audio Large Language Models (ALLMs), voice agents are now being deployed in high-stakes domains such as banking, customer service, and IT support. However, their vulnerabilities to adversarial misuse still remain unexplored. While prior work has examined aspects of trustworthiness in ALLMs, such as harmful content generation and hallucination, systematic security evaluations of voice agents are still lacking. To address this gap, we propose Aegis, a red-teaming framework for the governance, integrity, and security of voice agents. Aegis models the realistic deployment pipeline of voice agents and designs structured adversarial scenarios of critical risks, including privacy leakage, privilege escalation, resource abuse, etc. We evaluate the framework through case studies in banking call centers, IT Support, and logistics. Our evaluation shows that while access controls mitigate data-level risks, voice agents remain vulnerable to behavioral attacks that cannot be addressed through access restrictions alone, even under strict access controls. We observe systematic differences across model families, with open-weight models exhibiting higher susceptibility, underscoring the need for layered defenses that combine access control, policy enforcement, and behavioral monitoring to secure next-generation voice agents.

Activation steering is a practical post-training model alignment technique to enhance the utility of Large Language Models (LLMs). Prior to deploying a model as a service, developers can steer a pre-trained model toward specific behavioral objectives, such as compliance or instruction adherence, without the need for retraining. This process is as simple as adding a steering vector to the model's internal representations. However, this capability unintentionally introduces critical and under-explored safety risks. We identify a phenomenon termed Steering Externalities, where steering vectors derived from entirely benign datasets-such as those enforcing strict compliance or specific output formats like JSON-inadvertently erode safety guardrails. Experiments reveal that these interventions act as a force multiplier, creating new vulnerabilities to jailbreaks and increasing attack success rates to over 80% on standard benchmarks by bypassing the initial safety alignment. Ultimately, our results expose a critical blind spot in deployment: benign activation steering systematically erodes the "safety margin," rendering models more vulnerable to black-box attacks and proving that inference-time utility improvements must be rigorously audited for unintended safety externalities.

We present an experimental evaluation that assesses the robustness of four open source LLMs claiming function-calling capabilities against three different attacks, and we measure the effectiveness of eight different defences. Our results show how these models are not safe by default, and how the defences are not yet employable in real-world scenarios.

Membership inference attacks (MIAs) pose a critical privacy threat to fine-tuned large language models (LLMs), especially when models are adapted to domain-specific tasks using sensitive data. While prior black-box MIA techniques rely on confidence scores or token likelihoods, these signals are often entangled with a sample's intrinsic properties - such as content difficulty or rarity - leading to poor generalization and low signal-to-noise ratios. In this paper, we propose ICP-MIA, a novel MIA framework grounded in the theory of training dynamics, particularly the phenomenon of diminishing returns during optimization. We introduce the Optimization Gap as a fundamental signal of membership: at convergence, member samples exhibit minimal remaining loss-reduction potential, while non-members retain significant potential for further optimization. To estimate this gap in a black-box setting, we propose In-Context Probing (ICP), a training-free method that simulates fine-tuning-like behavior via strategically constructed input contexts. We propose two probing strategies: reference-data-based (using semantically similar public samples) and self-perturbation (via masking or generation). Experiments on three tasks and multiple LLMs show that ICP-MIA significantly outperforms prior black-box MIAs, particularly at low false positive rates. We further analyze how reference data alignment, model type, PEFT configurations, and training schedules affect attack effectiveness. Our findings establish ICP-MIA as a practical and theoretically grounded framework for auditing privacy risks in deployed LLMs.

Tool calling agents are an emerging paradigm in LLM deployment, with major platforms such as ChatGPT, Claude, and Gemini adding connectors and autonomous capabilities. However, the inherent unreliability of LLMs introduces fundamental security risks when these agents operate over sensitive user services. Prior approaches either rely on manually written policies that require security expertise, or place LLMs in the confinement loop, which lacks rigorous security guarantees. We present MiniScope, a framework that enables tool calling agents to operate on user accounts while confining potential damage from unreliable LLMs. MiniScope introduces a novel way to automatically and rigorously enforce least privilege principles by reconstructing permission hierarchies that reflect relationships among tool calls and combining them with a mobile-style permission model to balance security and ease of use. To evaluate MiniScope, we create a synthetic dataset derived from ten popular real-world applications, capturing the complexity of realistic agentic tasks beyond existing simplified benchmarks. Our evaluation shows that MiniScope incurs only 1-6% latency overhead compared to vanilla tool calling agents, while significantly outperforming the LLM based baseline in minimizing permissions as well as computational and operational costs.

Large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Existing approaches overwhelmingly operate within the prompt-optimization paradigm: whether through traditional algorithmic search or recent agent-based workflows, the resulting prompts typically retain malicious semantic signals that modern guardrails are primed to detect. In contrast, we identify a deeper, largely overlooked vulnerability stemming from the highly interconnected nature of an LLM's internal knowledge. This structure allows harmful objectives to be realized by weaving together sequences of benign sub-queries, each of which individually evades detection. To exploit this loophole, we introduce the Correlated Knowledge Attack Agent (CKA-Agent), a dynamic framework that reframes jailbreaking as an adaptive, tree-structured exploration of the target model's knowledge base. The CKA-Agent issues locally innocuous queries, uses model responses to guide exploration across multiple paths, and ultimately assembles the aggregated information to achieve the original harmful objective. Evaluated across state-of-the-art commercial LLMs (Gemini2.5-Flash/Pro, GPT-oss-120B, Claude-Haiku-4.5), CKA-Agent consistently achieves over 95% success rates even against strong guardrails, underscoring the severity of this vulnerability and the urgent need for defenses against such knowledge-decomposition attacks. Our codes are available at https://github.com/Graph-COM/CKA-Agent.

This work aims to investigate how different Large Language Models (LLMs) alignment methods affect the models' responses to prompt attacks. We selected open source models based on the most common alignment methods, namely, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Human Feedback (RLHF). We conducted a systematic analysis using statistical methods to verify how sensitive the Attack Success Rate (ASR) is when we apply variations to prompts designed to elicit inappropriate content from LLMs. Our results show that even small prompt modifications can significantly change the Attack Success Rate (ASR) according to the statistical tests we run, making the models more or less susceptible to types of attack. Critically, our results demonstrate that running existing 'attack benchmarks' alone may not be sufficient to elicit all possible vulnerabilities of both models and alignment methods. This paper thus contributes to ongoing efforts on model attack evaluation by means of systematic and statistically-based analyses of the different alignment methods and how sensitive their ASR is to prompt variation.

Unlearning in large language models (LLMs) is critical for regulatory compliance and for building ethical generative AI systems that avoid producing private, toxic, illegal, or copyrighted content. Despite rapid progress, in this work we show that \textit{almost all} existing unlearning methods fail to achieve true forgetting in practice. Specifically, while evaluations of these `unlearned' models under deterministic (greedy) decoding often suggest successful knowledge removal using standard benchmarks (as has been done in the literature), we show that sensitive information reliably resurfaces when models are sampled with standard probabilistic decoding. To rigorously capture this vulnerability, we introduce \texttt{leak@$k$}, a new meta-evaluation metric that quantifies the likelihood of forgotten knowledge reappearing when generating $k$ samples from the model under realistic decoding strategies. Using three widely adopted benchmarks, TOFU, MUSE, and WMDP, we conduct the first large-scale, systematic study of unlearning reliability using our newly defined \texttt{leak@$k$} metric. Our findings demonstrate that knowledge leakage persists across methods and tasks, underscoring that current state-of-the-art unlearning techniques provide only limited forgetting and highlighting the urgent need for more robust approaches to LLM unlearning. We propose an algorithm, termed Robust Unlearning under LEak@$k$ metric (\texttt{RULE}), which serves as an initial step toward addressing this concern. We demonstrate that \texttt{RULE} provides an unlearned model for TOFU benchmark with no information leakage for a large number of generation samples.

Large language model (LLM) unlearning has become a critical mechanism for removing undesired data, knowledge, or behaviors from pre-trained models while retaining their general utility. Yet, with the rise of open-weight LLMs, we ask: can the unlearning process itself be backdoored, appearing successful under normal conditions yet reverting to pre-unlearned behavior when a hidden trigger is activated? Drawing inspiration from classical backdoor attacks that embed triggers into training data to enforce specific behaviors, we investigate backdoor unlearning, where models forget as intended in the clean setting but recover forgotten knowledge when the trigger appears. We show that designing such attacks presents unique challenges, hinging on where triggers are placed and how backdoor training is reinforced. We uncover a strong link between backdoor efficacy and the attention sink phenomenon, i.e., shallow input tokens consistently attract disproportionate attention in LLMs. Our analysis reveals that these attention sinks serve as gateways for backdoor unlearning: placing triggers at sink positions and aligning their attention values markedly enhances backdoor persistence. Extensive experiments validate these findings, showing that attention-sink-guided backdoor unlearning reliably restores forgotten knowledge in the presence of backdoor triggers, while behaving indistinguishably from a normally unlearned model when triggers are absent. Code is available at https://github.com/OPTML-Group/Unlearn-Backdoor.

While LLM agents can plan multi-step tasks, intervening at the planning stage-before any action is executed-is often the safest way to prevent harm, since certain risks can lead to severe consequences once carried out. However, existing guardrails mostly operate post-execution, which is difficult to scale and leaves little room for controllable supervision at the plan level. To address this challenge, we highlight three critical gaps in current research: data gap, model gap, and evaluation gap. To close the data gap, we introduce AuraGen, a controllable engine that (i) synthesizes benign trajectories, (ii) injects category-labeled risks with calibrated difficulty, and (iii) filters outputs via an automated reward model, producing large and reliable corpora for pre-execution safety. To close the guardian model gap, we propose a foundational guardrail Safiron, combining a cross-planner adapter with a compact guardian model. The adapter unifies different input formats, while Safiron flags risky cases, assigns risk types, and generates rationales; trained in two stages with a broadly explored data recipe, Safiron achieves robust transfer across settings. To close the evaluation gap, we release Pre-Exec Bench, a realistic benchmark covering diverse tools and branching trajectories, which measures detection, fine-grained categorization, explanation, and cross-planner generalization in human-verified scenarios. Extensive experiments demonstrate consistent gains of the proposed guardrail over strong baselines on Pre-Exec Bench, and ablations further distill actionable practices, providing a practical template for safer agentic systems.

Machine unlearning for large language models (LLMs) aims to remove undesired data, knowledge, and behaviors (e.g., for safety, privacy, or copyright) while preserving useful model capabilities. Despite rapid progress over the past two years, research in LLM unlearning remains fragmented, with limited clarity on what constitutes effective unlearning and how it should be rigorously evaluated. In this work, we present a principled taxonomy of twelve recent stateful unlearning methods, grouped into three methodological families: divergence-driven optimization, representation misalignment, and rejection-based targeted unlearning. Building on this taxonomy, we revisit the evaluation of unlearning effectiveness (UE), utility retention (UT), and robustness (Rob), focusing on the WMDP benchmark. Our analysis shows that current evaluations, dominated by multiple-choice question (MCQ) accuracy, offer only a narrow perspective, often overstating success while overlooking the model's actual generation behavior. To address this gap, we introduce open question-answering (Open-QA) metrics that better capture generative performance and reveal the inherent UE-UT tradeoff across method families. Furthermore, we demonstrate that robustness requires finer-grained analysis: for example, vulnerabilities differ substantially between in-domain relearning and out-of-domain fine-tuning, even though both fall under model-level attacks. Through this study, we hope to deliver a full-stack revisit of LLM unlearning and actionable guidance for designing and evaluating future methods.

Large reasoning models (LRMs) "think" by generating structured chain-of-thought (CoT) before producing a final answer, yet they still lack the ability to reason critically about safety alignment and are easily biased when a flawed premise is injected into their thought process. We propose RECAP (Robust Safety Alignment via Counter-Aligned Prefilling), a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories and reroute to safe and helpful responses. RECAP trains on a mixture of synthetically generated counter-aligned CoT prefills and standard prompts, requires no additional training cost or modifications beyond vanilla reinforcement learning from human feedback (RLHF), and substantially improves safety and jailbreak robustness, reduces overrefusal, and preserves core reasoning capability -- all while maintaining inference token budget. Extensive analysis shows that RECAP-trained models engage in self-reflection more frequently and remain robust under adaptive attacks, preserving safety even after repeated attempts to override their reasoning.

Large language model (LLM) unlearning aims to surgically remove the influence of undesired data or knowledge from an existing model while preserving its utility on unrelated tasks. This paradigm has shown promise in addressing privacy and safety concerns. However, recent findings reveal that unlearning effects are often fragile: post-unlearning manipulations such as weight quantization or fine-tuning can quickly neutralize the intended forgetting. Prior efforts to improve robustness primarily reformulate unlearning objectives by explicitly assuming the role of vulnerability sources. In this work, we take a different perspective by investigating the role of the optimizer, independent of unlearning objectives and formulations, in shaping unlearning robustness. We show that the 'grade' of the optimizer, defined by the level of information it exploits, ranging from zeroth-order (gradient-free) to first-order (gradient-based) to second-order (Hessian-based), is tightly linked to the resilience of unlearning. Surprisingly, we find that downgrading the optimizer, such as using zeroth-order methods or compressed-gradient variants (e.g., gradient sign-based optimizers), often leads to stronger robustness. While these optimizers produce noisier and less precise updates, they encourage convergence to harder-to-disturb basins in the loss landscape, thereby resisting post-training perturbations. By connecting zeroth-order methods with randomized smoothing, we further highlight their natural advantage for robust unlearning. Motivated by these insights, we propose a hybrid optimizer that combines first-order and zeroth-order updates, preserving unlearning efficacy while enhancing robustness. Extensive experiments on the MUSE and WMDP benchmarks, across multiple LLM unlearning algorithms, validate that our approach achieves more resilient forgetting without sacrificing unlearning quality.

Detecting AI-generated text is an increasing necessity to combat misuse of LLMs in education, business compliance, journalism, and social media, where synthetic fluency can mask misinformation or deception. While prior detectors often rely on token-level likelihoods or opaque black-box classifiers, these approaches struggle against high-quality generations and offer little interpretability. In this work, we propose DivEye, a novel detection framework that captures how unpredictability fluctuates across a text using surprisal-based features. Motivated by the observation that human-authored text exhibits richer variability in lexical and structural unpredictability than LLM outputs, DivEye captures this signal through a set of interpretable statistical features. Our method outperforms existing zero-shot detectors by up to 33.2% and achieves competitive performance with fine-tuned baselines across multiple benchmarks. DivEye is robust to paraphrasing and adversarial attacks, generalizes well across domains and models, and improves the performance of existing detectors by up to 18.7% when used as an auxiliary signal. Beyond detection, DivEye provides interpretable insights into why a text is flagged, pointing to rhythmic unpredictability as a powerful and underexplored signal for LLM detection.

Model stealing attacks endanger the confidentiality of machine learning models offered as a service. Although these models are kept secret, a malicious party can query a model to label data samples and train their own substitute model, violating intellectual property. While novel attacks in the field are continually being published, their design and evaluations are not standardised, making it challenging to compare prior works and assess progress in the field. This paper is the first to address this gap by providing recommendations for designing and evaluating model stealing attacks. To this end, we study the largest group of attacks that rely on training a substitute model -- those attacking image classification models. We propose the first comprehensive threat model and develop a framework for attack comparison. Further, we analyse attack setups from related works to understand which tasks and models have been studied the most. Based on our findings, we present best practices for attack development before, during, and beyond experiments and derive an extensive list of open research questions regarding the evaluation of model stealing attacks. Our findings and recommendations also transfer to other problem domains, hence establishing the first generic evaluation methodology for model stealing attacks.