ICML 2026 Papers

Large Language Model (LLM) agents use memory to learn from past interactions. However, this reliance on memory introduces a critical security risk: an adversary can inject seemingly harmless records into an agent's memory to manipulate its future behavior. This vulnerability is characterized by two core aspects: First, the malicious effect of injected records is only activated within a specific context, making them hard to detect when individual memory entries are audited in isolation. Second, once triggered, the manipulation can initiate a self-reinforcing error cycle: the corrupted outcome is stored as precedent, which not only amplifies the initial error but also progressively lowers the threshold for similar attacks in the future. To address these challenges, we introduce \emph{A-MemGuard} (\underline{A}gent-\underline{Mem}ory \underline{Guard}), the first defense framework for LLM agent memory. The core idea of our work is the insight that memory itself must become both \emph{self-checking} and \emph{self-correcting}. Without modifying the agent's core architecture, A-MemGuard combines two mechanisms: (1) \textbf{consensus-based validation}, which detects anomalies by comparing reasoning paths derived from multiple related memories and (2) a \textbf{dual-memory structure}, where detected failures are distilled into ``lessons'' stored separately and consulted before future actions, breaking error cycles and enabling adaptation. Comprehensive evaluations on multiple benchmarks show that A-MemGuard effectively cuts attack success rates by over 95\% while incurring a minimal utility cost. This work shifts LLM memory security from static filtering to a proactive, experience-driven model where defenses strengthen over time.

Social interactions are characterized by both adversarial and cooperative aspects. Communications between agents may also involve adversarially motivated actors. Messages can pass through intermediaries with malicious intent before reaching the intended receiver. These actors may modify the message to induce misunderstanding from the receiver while preserving the overall characteristics of the message. This form of misinformation is prevalent in real-world communications and may affect the dynamics under which communication protocols are developed. However, this aspect of social interaction is relatively underexplored in many emergent communication studies that aim to understand the environmental factors behind the emergence of languages' characteristics. This work explores how misinformation affects language emergence with a focus on compositionality. We design a communication game containing a malign intermediary between the sender and receiver. We find that risks of malign misrepresentation promote the emergence of compositional languages in simulations of communicative agents. Furthermore, we observe that adaptability of malign intermediaries is a crucial factor in forming a consistent pressure toward compositionality and that partial misinformation, in which the intermediary targets only a subset of attributes, can also induce compositionality.

From generating headlines to fabricating news, the Large Language Models (LLMs) are typically assessed by their final outputs, under the safety assumption that a refusal response signifies safe reasoning throughout the entire process. Challenging this assumption, our study reveals that during fake news generation, even when a model rejects a harmful request, its Chain-of-Thought (CoT) reasoning may still internally contain and propagate unsafe narratives. To analyze this phenomenon, we introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates the role of individual attention heads through Jacobian-based spectral metrics. Within this framework, we introduce three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond or embed deceptive reasoning patterns. Extensive experiments on multiple reasoning-oriented LLMs show that the generation risk rise significantly when the thinking mode is activated, where the critical routing decisions concentrated in only a few contiguous mid-depth layers. By precisely identifying the attention heads responsible for this divergence, our work challenges the assumption that refusal implies safety and provides a new understanding perspective for mitigating latent reasoning risks.

As LLM-based agents increasingly operate in high-stakes domains with real-world consequences, ensuring their behavioral safety becomes paramount. The dominant oversight paradigm, *LLM-as-a-Judge*, faces a fundamental dilemma: how can probabilistic systems reliably supervise other probabilistic systems without inheriting their failure modes? We argue that formal verification offers a principled escape from this dilemma, yet its adoption has been hindered by a critical bottleneck: the translation from natural language requirements to formal specifications. This paper bridges this gap by proposing , a neuro-symbolic framework that employs a bidirectional **Formal-of-Thought** architecture: LLMs serve as *specification compilers* that top-down decompose high-level human intent into atomic, verifiable constraints, then bottom-up prove compliance using Dafny specifications and Z3 Satisfiability modulo theories solving, which produces mathematical guarantees rather than probabilistic scores. We validate across three benchmarks spanning behavioral safety, multi-domain constraint adherence, and agentic upward deception detection. Experiments on 7 agent models demonstrate that achieves an average improvement of **16.6%** over LLM-as-a-Judge baselines, enables *weak-to-strong* generalization where a 7B judge achieves over 90% accuracy detecting deception from 72B agents, and provides *near-linear safety improvement* through iterative refinement.

Agentic scaffolds have dramatically improved LLM performance on complex, long-horizon tasks, yielding both broad benefits and amplified risks in domains like cybersecurity. Existing benchmarks for AI agents in cybersecurity focus mainly on measuring proficiency–how effectively agents can complete offensive security tasks–but neglect a critical question: when and how should agents refuse harmful requests? We present the first framework for establishing refusal boundaries in offensive security contexts. Our framework defines (1) principled criteria for when tasks should be refused, (2) categories of tasks that warrant refusal, and (3) evaluation methodology for measuring agent robustness under both benign and adversarial conditions. We apply this framework to assess how current LLM-powered agents adhere to appropriate refusal boundaries across a range of web-based offensive security scenarios, finding that 6 of 8 frontier models tested refuse to identify vulnerabilities in real systems 0\% of the time.

The limited transferability of adversarial attacks on Vision-Language Models (VLMs) stems from their failure to navigate model-specific safety alignments, where superficial perturbations exploit surrogate-specific artifacts rather than shared safety-critical features. We reveal through linear probing that safety-related representations are concentrated within specific intermediate neuronal circuits, which act as localized defense bottlenecks that can be disentangled from transferable features. To overcome this barrier, we propose the Safety Circuit Intervention Attack (SCIA), a framework that surgically steers internal representations to bypass these localized safety mechanisms. SCIA employs a dual-objective steering strategy that suppresses the defensive circuit encoding safety features while amplifying the transferable circuit capturing model-agnostic representations, effectively decoupling adversarial patterns from surrogate-specific safety behaviors. Furthermore, we incorporate contrastive semantic steering and spectral smoothness regularization to guide optimization toward compliant semantic regions while producing visually coherent perturbations. Experimental results demonstrate that SCIA significantly outperforms state-of-the-art methods in bypassing unseen black-box VLMs.

Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we introduce and formalize SAE ensembles. Furthermore, we propose to ensemble multiple SAEs through *naive bagging* and *boosting*. In naive bagging, SAEs trained with different weight initializations are ensembled, whereas in boosting SAEs sequentially trained to minimize the residual error are ensembled. Theoretically, naive bagging and boosting are justified as approaches to reduce reconstruction error. Empirically, we evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that, compared to an expanded SAE that matches the number of features in the ensemble, ensembling SAEs improves the reconstruction of language model activations along with SAE stability. Additionally, on downstream tasks such as concept detection and spurious correlation removal, SAE ensembles achieve better performance, showing improved practical utility.

Large Language Models (LLMs) often provide chain-of-thought (CoT) reasoning traces that appear plausible, but may hide internal biases. We call these *unverbalized biases*. Monitoring models via their stated reasoning is therefore unreliable, and existing bias evaluations typically require predefined categories and hand-crafted datasets. In this work, we introduce a fully automated, black-box pipeline for detecting task-specific unverbalized biases. Given a task dataset, the pipeline uses LLM autoraters to generate candidate bias concepts. It then tests each concept on progressively larger input samples by generating positive and negative variations, and applies statistical techniques for multiple testing and early stopping. A concept is flagged as an unverbalized bias if it yields statistically significant performance differences while not being cited as justification in the model's CoTs. We evaluate our pipeline across seven LLMs on three decision tasks (hiring, loan approval, and university admissions). Our technique automatically discovers previously unknown biases in these models (e.g., Spanish fluency, English proficiency, writing formality). In the same run, the pipeline also validates biases that were manually identified by prior work (gender, race, religion, ethnicity). More broadly, our proposed approach provides a practical, scalable path to automatic, more efficient, and broader task-specific unverbalized bias discovery.

Despite the central role of attention heads in Transformers, we lack tools to understand why a model attends to a particular token. To address this, we study the query-key (QK) space -- the bilinear joint embedding space between queries and keys. We present a contrastive covariance method to decompose the QK space into low-rank, human-interpretable components. It is when features in keys and queries align in these low-rank subspaces that high attention scores are produced. We first study our method both analytically and empirically in a simplified setting. We then apply our method to large language models to identify human-interpretable QK subspaces for categorical semantic features and binding features. Finally, we demonstrate how attention scores can be attributed to our identified features.

Backdoor and data-poisoning attacks can achieve high attack success while evading existing spectral and optimisation-based defences. We show that this behaviour is not incidental, but arises from a fundamental geometric mechanism in input space. Using kernel ridge regression as an exact model of wide neural networks, we prove that clustered dirty-label poisons induce a rank-one spike in the input Hessian whose magnitude scales quadratically with attack efficacy. Crucially, for nonlinear kernels we identify a near-clone regime in which poison efficacy remains order-one while the induced input curvature vanishes, making the attack provably spectrally undetectable. We further show that input-gradient regularisation contracts poison-aligned Fisher and Hessian eigenmodes under gradient flow, yielding an explicit and unavoidable safety–efficacy trade-off by reducing data-fitting capacity. For exponential kernels, this defence admits a precise interpretation as an anisotropic high-pass filter that increases the effective length scale and suppresses near-clone poisons. Extensive experiments on linear models and deep convolutional networks across MNIST and CIFAR-10/100 validate the theory, demonstrating consistent lags between attack success and spectral visibility, and showing that regularisation and data augmentation jointly suppress poisoning. Our results establish when backdoors are inherently invisible, and provide the first end-to-end characterisation of poisoning, detectability, and defence through input-space curvature.

The standard way to evaluate language models on subjective tasks is through pairwise comparisons: an annotator chooses the "better" of two model responses for a given prompt. These comparisons are then aggregated into a single ranking via the Bradley–Terry (BT) framework, forcing heterogeneous preferences into a total order and violating basic social-choice desiderata. In contrast, social choice theory provides an alternative approach called maximal lotteries, which aggregates pairwise preferences without imposing any assumptions on their structure. However, we show that maximal lotteries can be highly sensitive to heterogeneity among annotators and across prompts. We introduce *robust lotteries*, which optimize worst-case performance under plausible shifts in the preference data. On large-scale preference datasets, robust lotteries achieve more reliable win rate guarantees across the annotator distribution and recover a stable set of top performing models.

Safety alignment in large language models (LLMs) and multimodal large language models (MLLMs) is commonly assumed to operate as a near-binary threshold mechanism. We challenge this assumption by revealing that safety behavior is governed by an \emph{instability region} where small perturbations induce stochastic refusal decisions rather than deterministic outcomes. We develop a multi-metric diagnostic framework combining external and internal signals to characterize this instability. Through systematic experiments, we identify a characteristic \emph{diagnostic signature}: inputs in unstable regimes exhibit elevated output uncertainty yet \emph{decreased} internal safety activation, a decoupling phenomenon that explains why detection-based defenses fail against sophisticated attacks. Building on this framework, we introduce \textbf{Furina}, a jailbreak attack that deliberately induces this signature through fragmented, scene-anchored prompts without model-specific optimization. Furina outperforms strong single-turn and multi-turn baselines on HarmBench and achieves competitive results on MM-SafetyBench, demonstrating that uncertainty amplification provides a principled and transferable mechanism for understanding safety vulnerabilities. Code is available at: \url{https://github.com/0xCavaliers/Furina_Jailbreak}.

We develop a theory of intelligent agency grounded in probabilistic modeling for neural models. Agents are represented as outcome distributions with epistemic utility given by log score, and compositions are defined through weighted logarithmic pooling that strictly improves every member's welfare. We prove that strict unanimity is impossible under linear pooling or in binary outcome spaces, but possible with three or more outcomes. Our framework admits recursive structure via cloning invariance, continuity, and openness, while tilt-based analysis rules out trivial duplication. Finally, we formalize an agentic alignment phenomenon in LLMs using our theory: eliciting a benevolent persona ("Luigi'') induces an antagonistic counterpart ("Waluigi''), while a manifest-then-suppress Waluigi strategy yields strictly larger first-order misalignment reduction than pure Luigi reinforcement alone. These results clarify how developing a principled mathematical framework for how subagents can coalesce into coherent higher-level entities provides novel implications for alignment in agentic AI systems.

Retrieval-augmented generation (RAG) systems are vulnerable to attacks that inject poisoned passages into the retrieved context, even at low corruption rates. We show that existing attacks are not designed to be stealthy, allowing reliable detection and mitigation. We formalize a distinguishability-based security game to quantify stealth for such attacks. If a few poisoned passages control the response, they must bias the inference process more than the benign ones, inherently compromising stealth. This motivates analyzing intermediate signals of LLMs, such as attention weights, to approximate the influence of different passages on the response. Leveraging attention weights, we introduce the Normalized Passage Attention Score (NPAS) and a lightweight Attention-Variance Filter (AV Filter) that flags anomalous passages. Our method improves robustness, yielding up to ~20% higher accuracy than baseline defenses. We also develop adaptive attacks that attempt to conceal such anomalies, achieving up to 35% success rate and underscoring the challenges of achieving true stealth in poisoning RAG systems.

Machine unlearning seeks to remove the influence of specified data from a trained model. While the unlearning accuracy provides a widely used metric for assessing unlearning performance, it falls short in assessing the reliability of forgetting. In this paper, we find that the forgetting data points misclassified by unlearning accuracy still have their ground truth labels included in the conformal prediction set from the uncertainty quantification perspective, leading to a phenomenon we term fake forgetting. To address this issue, we propose a novel metric CR, inspired by conformal prediction, that offers a more reliable assessment of forgetting quality. Building on these insights, we further propose an unlearning framework CPU that incorporates conformal prediction into the Carlini & Wagner adversarial attack loss, enabling the ground truth label to be effectively removed from the conformal prediction set. Through extensive experiments on image classification tasks, we demonstrate both the effectiveness of our proposed metric and the superior forgetting quality achieved by our framework. Code is available at https://github.com/TIML-Group/Conformal-Prediction-Unlearning.

Large language models (LLMs) are pretrained on corpora containing trillions of tokens and, therefore, inevitably memorize sensitive information. Locate-then-edit methods, as a mainstream paradigm of model editing, offer a promising solution by modifying model parameters without retraining. However, in this work, we reveal a critical vulnerability of this paradigm: the parameter updates inadvertently serve as a side channel, enabling attackers to recover the edited data. We propose a two-stage reverse-engineering attack named KSTER (KeySpaceReconsTruction-then-EntropyReduction) that leverages the low-rank structure of these updates. First, we theoretically show that the row space of the update matrix encodes a "fingerprint" of the edited subjects, enabling accurate subject recovery via spectral analysis. Second, we introduce an entropy-based prompt recovery attack that reconstructs the semantic context of the edit. Extensive experiments on multiple LLMs demonstrate that our attacks can recover edited data with high success rates. Furthermore, we propose subspace camouflage, a defense strategy that obfuscates the update fingerprint with semantic decoys. This approach effectively mitigates reconstruction risks without compromising editing utility. Our code is available at https://github.com/reanatom/EditingAttack.

Existing data poisoning attacks on retrieval-augmented generation (RAG) systems scale poorly because they require costly optimization of poisoned documents for each target phrase. We introduce Eyes-on-Me, a modular attack that decomposes an adversarial document into reusable **Attention Attractors** and **Focus Regions**. Attractors are optimized to direct attention to the Focus Region. Attackers can then insert semantic baits for the retriever or malicious instructions for the generator, adapting to new targets at near zero cost. This is achieved by steering a small subset of attention heads that we empirically identify as strongly correlated with attack success. Across 18 end-to-end RAG settings (3 datasets $\times$ 2 retrievers $\times$ 3 generators), Eyes-on-Me raises average attack success rates from 21.9 to 57.8 (+35.9 points, 2.6$\times$ over prior work). A single optimized attractor transfers to unseen black box retrievers and generators without retraining. Our findings establish a scalable paradigm for RAG data poisoning and show that modular, reusable components pose a practical threat to modern AI systems. They also reveal a strong link between attention concentration and model outputs, informing interpretability research.

Activation steering offers a computationally efficient mechanism for controlling Large Language Models (LLMs) without fine-tuning. While effectively controlling target traits (e.g., persona), coherency degradation remains a major obstacle to safety and practical deployment. We hypothesize that this degradation stems from intervening on the residual stream, which indiscriminately affects aggregated features and inadvertently amplifies off-target noise. In this work, we identify a sparse subset of attention heads (only three heads) that independently govern persona and style formation, which we term *Style Modulation Heads*. Specifically, these heads can be localized via geometric analysis of internal representations, combining layer-wise cosine similarity and head-wise contribution scores. We demonstrate that intervention targeting only these specific heads achieves robust behavioral control while significantly mitigating the coherency degradation observed in residual stream steering. More broadly, our findings show that precise, component-level localization enables safer and more precise model control.

Conformal prediction (CP) has become a cornerstone of distribution-free uncertainty quantification, conventionally evaluated by its coverage and interval length. This work critically examines the sufficiency of these standard metrics. We demonstrate that the interval length might be deceptively improved through a counter-intuitive approach termed Prejudicial Trick (PT), while the coverage remains valid. Specifically, for any given test sample, PT probabilistically returns an interval, which is either null or constructed using an adjusted confidence level, thereby preserving marginal coverage. While PT potentially yields a deceptively lower interval length, it introduces practical vulnerabilities: the same input can yield completely different prediction intervals across repeated runs of the algorithm. We formally derive the conditions under which PT achieves these misleading improvements and provide extensive empirical evidence across various regression and classification tasks. Furthermore, we introduce a new metric interval stability which helps detect whether a new CP method implicitly improves the length based on such PT-like techniques. Code is available at https://github.com/benben-cd/PT-Conformal-Prediction