Bad

Regional linguistic variation remains an important challenge for Bangla natural language processing, particularly in informal and non-standard text. This paper introduces BanglaDial-Abuse, a balanced Bengali-script dataset developed for regional dialect identification in abusive and hostile Bangla text. The dataset contains 1,000 sentences distributed equally across four linguistic varieties: Standard Bangla, Chattagram, Sylhet, and Barishal, with 250 samples per class. The resource was constructed using a corpus-grounded synthetic procedure incorporating regional variation in pronouns, possessive forms, verb morphology, negation, interrogative structures, postpositions, vocabulary, and Bengali-script spelling conventions while preserving the underlying hostile or abusive meaning. Descriptive analysis shows broadly comparable sentence-length distributions but partially distinct lexical spaces across the four classes. Pairwise Jaccard vocabulary similarity ranges from 0.37 to 0.56. The primary task is four-class regional dialect identification rather than binary abusive-text detection. The dataset is publicly available through Zenodo under a Creative Commons Attribution 4.0 license. The current version is intended as a research and prototyping corpus rather than a native-speaker-validated gold-standard linguistic resource. Keywords: Bangla, Bengali, dialect identification, regional dialect, abusive language, low-resource NLP, Chattagram, Sylhet, Barishal, dataset

Interactive video generation (IVG) models have achieved remarkable progress in producing controllable visual content guided by user-defined actions, yet their security vulnerabilities remain largely unexplored. In this paper, we present the first systematic study of backdoor attacks against the interactivity of IVG models. Based on this attack surface, we propose BadAction, which leverages action-guided triggers to achieve the attack. Specifically, BadAction implants predefined motion patterns into the action sequences of backdoor samples and associates them with a static target video. Once triggered, the backdoored model generates frozen future frames that no longer respond to subsequent user actions, while preserving normal behavior on benign action sequences. In addition, we explore a stealthier attack in which multimodal triggers jointly poison action, text, and image inputs. Experiments show that BadAction achieves average attack success rates of 91.0% with action-only triggers and 80.4% with multimodal triggers. Moreover, extensive defense evaluations show that BadAction successfully bypasses existing backdoor detection methods, revealing a critical security gap in the interactive video generation pipeline. Project page: https://wsad55.github.io/badaction01/.

Fine-tuning large language models on narrow, misaligned tasks can undo their post-training alignment and induce novel misaligned behaviors -- a phenomenon known as \emph{emergent misalignment} (EM). EM has been linked to persona-like representations, where fine-tuning might reduce loss by amplifying a harmful or 'evil' persona. It remains unclear which properties of the training data drive this effect: whether all harmful examples contribute approximately equally to misalignment and whether different models are equally affected by the same fine-tuning examples. In this work, we use training data attribution to quantitatively estimate how much each harmful example contributes to EM. We benchmark the quality of the attribution via retraining -- a sound attribution score should enable us to enhance or attenuate EM by filtering data on that score. Score-based filtering can substantially enhance or attenuate EM; we find that both data-attribution scores and a black-box harmfulness score can identify consequential examples. All models we test become misaligned when trained on the same dataset, and influence scores perform best when filtering data from the same model that computed them. We find cross-model generalization of influence scores from scores derived from the three model families we tested, but this generalization does not recover same model filtering performance.

Machine unlearning has emerged as a critical post-hoc safety measure to erase sensitive concepts from Text-to-Image (T2I) models without prohibitive retraining. However, we reveal that current state-of-the-art (SOTA) approaches are brittle due to a severe lack of robustness to noise initialization. We call this phenomenon ``probabilistic forgetting'': suppressed concepts re-emerge under specific random initial noise conditions, despite appearing unlearned on other initializations. We trace this failure to the misalignment between standard Gaussian sampling during unlearning and the unlearning objective. Since the target concept manifests only in specific initial noise regions throughout the unlearning phase, uniform random sampling yields sparse, uninformative gradient updates that fail to drive robust erasure. To overcome this issue, we propose an adaptive, concept-conditioned sampling strategy that dynamically concentrates gradient updates on regions where the target concept manifests, down-weighting uninformative areas. We integrate our framework with six distinct SOTA unlearning methods across four diffusion backbones and evaluate it across safety, object, and artistic-style unlearning, as well as under black-box and white-box adversarial attacks. Our method reduces the conditional nudity re-emergence rate across random initializations by 67.2% on average over four baselines and lowers attack success rates across both adversarial evaluations. Across concept domains, Adaptive Noise Sampling strengthens adversarial robustness and non-target retention while preserving competitive generative quality and target-erasure performance.

Effective communication between users and AI agents is essential for human-AI collaboration. The XY problem is a well-known communication pitfall where a person asks about their attempted solution rather than their actual problem. We extend prior sycophancy evaluation to the XY problem in agentic settings, evaluating whether agents can resist plausible but misleading suggestions from users and communicate their reasoning. We introduce XYEval, a meta-evaluation framework that can transform an existing benchmark into an XY problem evaluation. We evaluate five models across six diverse benchmark suites. Agents suffer large XY drops under XY mutation across benchmarks, with relative drops reaching up to 46.7%. With $τ^2$-bench, we further show that agent performance drops more when encountering a pedantic user who requires detailed explanations before approving a better solution. Our findings suggest that current agents lack the ability to effectively reason and communicate when facing misleading suggestions. A simple system instruction baseline that encourages awareness of XY problems only offers partial mitigation. Extensive trace analyses provide behavioral insights into how and why these XY drops occur across execution trajectories. Our results show that mitigating the XY problem remains challenging, requiring agents to both recognize user misdirection and clearly communicate the underlying problem.

Reliable agent evaluation is complicated by automatic harness optimization, which repeatedly uses a released benchmark $B_{\mathrm{rel}}$ to guide a Proposer that edits prompts, memory, retrieval, tools, and control code around a fixed foundation model. Task holdout is commonly used to guard against harness overfitting. It varies semantic tasks but leaves the benchmark protocol fixed, so a bad genius Proposer can produce a cheating harness whose improvement over the initial harness on $B_{\mathrm{rel}}$ depends on a benchmark-wide shortcut. We introduce Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over valid counterfactual benchmarks. After each Proposer update, a Challenger searches for an executable protocol transformation with large gain destruction. A validity firewall checks that task semantics are preserved, while a held-out confirmation set determines whether the counterfactual enters a finite archive. We formalize an ideal shortcut-neutralized benchmark $B_0$ and establish theoretical guarantees linking finite counterfactual archives to $B_0$ and characterizing sequential Challenger search. We evaluate CHASE on Syn-Ledger and OfficeQA, where CHASE retains strong released-benchmark gains while substantially reducing gain destruction under valid protocol transformations.

Robust comparisons of federated aggregation methods require joint consideration of predictive performance, threat definitions, metric semantics, and execution provenance. A 500-cell seed-1 evaluation matrix was reconstructed across five aggregation methods, five datasets, five architectures, and four recorded conditions: clean, sign-flipping, Gaussian, and BadNets. Successful execution logs were identified for 454 original runs and 36 repaired or rerun executions, whereas 10 clean SVHN cells were supported by summary-only provenance. Trimmed Mean achieved the highest clean macro-mean accuracy (76.02%) and the lowest mean within-task rank (1.70). Krum attained the highest recorded accuracy under both sign-flipping and Gaussian configurations. These relative rankings remained unchanged when analysis was restricted to 21 task pairs for which original successful logs were available for every method-condition combination. Audit of the supplied BadNets metric implementation established that every test input is triggered prior to target-label counting; consequently, the retained metric represents Triggered Target-Label Rate (TTLR) rather than a conventional target-excluding attack success rate. An audit of the supplied FedPARETO scaffold further identified a pathway in which predictive summaries may characterize an uncorrupted local model while the aggregation weight is applied to a separately corrupted update, introducing a potential discrepancy between reported predictive outcomes and the updates used for aggregation. The canonical matrix contains a single identified seed for each cell, and exact attack and configuration lineage is incomplete. Accordingly, the findings should be interpreted as descriptive comparisons within the recorded configurations and not as statistical estimates or universal claims regarding robustness.

Good action rankings do not make a contrastive critic safe to maximize. These critics increasingly act as value-like objectives for best-of-$K$ selection, planning, and critic-guided generation. Unbounded bilinear scores can let large embedding norms inflate off-support values, but cosine bounding does not remove the failure. A controlled support decomposition attributes most raw bilinear regret to norm drift. Cosine and hybrid critics nevertheless select off-support actions from most pools and incur comparable regret. Contrastive scores are weakly calibrated or inverted in the top score decile across four OGBench navigation tasks, and they fail to order fixed-query actions by value. Bellman-trained TD-Q succeeds, including in a parameter-matched function-class control. Realized costs depend on the task: simulator rollouts reveal single-step selection costs on PointMaze and the exact-$Q^*$ toy but well-powered nulls on AntMaze and HumanoidMaze, where the controller can self-correct. A training/readout decomposition traces the lost ordering to the cosine training objective; raw-trained embeddings retain weak ordering after inference-time normalization. Candidate maximization can therefore exploit false positives caused by norm drift, score saturation, or in-support misranking. Contrastive critics remain useful compatibility rankers on navigation and manipulation tasks, but action selection requires a value-calibrated scalar.

World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce BadWAM, a unified framework for modeling and evaluating World-Action Drift Attacks: a new class of WAM-specific adversarial attacks that use small visual perturbations to break the alignment between what a WAM imagines and what it executes. BadWAM characterizes this attack surface along two natural criteria: attack strength and stealthiness. When the adversary prioritizes disruption, BadWAM instantiates an action-only adversarial attack, which directly drives the model toward task-failing actions. When the adversary additionally prioritizes stealth, BadWAM instantiates an imagination-preserving adversarial attack, which seeks to induce harmful action shifts while keeping the model's predicted future close to its clean imagination. Together, these two attacks capture a spectrum of WAM-specific failures: from overt action hijacking to stealthier cases where the model appears to imagine a plausible future but executes a desynchronized action. We evaluate BadWAM across different variants of WAMs. Results show that our attacks substantially reduce task success rates under closed-loop execution. For example, our action-only attack reduces the model performance from 96.5% to 43.1% success. The results of our imagination-preserving attack further exposes a WAM-specific vulnerability: moderate future-preserving regularization can maintain strong attack performance while reducing future imagination drift.

How robust and generalisable are pathology foundation models and have their scaling limites been reached? We benchmarked twelve pathology foundation models (PFMs) and ResNet baselines using our Robustness Evaluation and Enhancement Toolbox (REET) across eleven clinically realistic perturbations and a dissimilarity-driven Non-Redundant K-fold validation (NR-Kfold) protocol. We introduce a Perturbation Performance Index (PPI) to summarise accuracy trends under controlled perturbation sweeps and analyse robustness scaling with parameter count. We show that PFMs consistently outperform CNNs in both robustness and domain generalisation, yet model scaling shows diminishing returns: mid-sized models such (UNI2/Virchow-2 etc.) achieve comparable or greater resilience than larger systems. NR-Kfold analysis further reveals systematic accuracy loss and increased variability when training-test similarity is broken, underscoring the need for explicit distribution-shift evaluation. These findings suggest that the next generation of pathology foundation models must prioritise data quality, multimodality information and domain alignment over parameter count to achieve genuine clinical reliability.

The game of Taboo requires describing a target word without using a set of forbidden words, so that other players can guess it. This deceptively simple task combines strict lexical constraints with the need for communicatively effective descriptions, making it a compelling playground for examining how LLMs navigate competing demands at inference time. We evaluate two open-weight models under conditions that intervene at progressively deeper levels of the generative process, from prompting to generation-time constraints to internal representations manipulations. We assess their outputs through forbidden word violation detection, LLM-as-a-judge measuring the degree to which generated descriptions successfully evoke the target concept for both human and machine guessers, and examining whether the strategies models adopt under constraint align with those of human players. Our results show that compliance with the rules of the game and communicative effectiveness trade off differently across conditions, and that models remain substantially weaker than humans as guessers, suggesting that lexical grounding under constraint is an open challenge for current language models.

While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." Unintentionally bad contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.

Visual world models (VWMs) synthesize interactive, action-conditioned rollouts from a single context image. However, it remains an open question how robust these models are to adversarial perturbations. Standard adversarial attacks fail to assess this vulnerability because attackers lack ground-truth future videos and cannot predict subsequent user controls. We introduce BadWorld, a label-free adversarial framework tailored for autoregressive VWMs that systematically overcomes both constraints. First, to bypass the need for future supervision, we propose a self-supervised velocity attack that directly disrupts the early denoising dynamics of the model. Second, to ensure the attack generalizes across unpredictable user actions, we formulate a trajectory-adaptive bi-level optimization that actively mines hard control sequences to forge control-agnostic perturbations. Evaluated on representative VWMs with continuous and discrete controls, BadWorld exposes severe structural fragility. Visually indistinguishable adversarial images reliably trigger catastrophic degradation in future rollouts, leading to incomplete denoising, structural collapse, and control inconsistency. These findings reveal critical risks for deploying VWMs in safety-critical systems while highlighting a practical mechanism for privacy protection.

This position paper argues that to obtain reliable embodied reward models, the community must invest in ``bad'' robot data: failed, suboptimal, error-prone, and even hazardous behaviors. While reward models are central to any foundation model's lifecycle, today's embodied reward models are trained primarily on successful behaviors. We analyze three state-of-the-art embodied reward models and find that they systematically over-reward behaviors that real human evaluators would penalize, including unsafe interactions, poor execution, and shortcut strategies that only superficially satisfy tasks. We attribute these failures to a key data gap: the scarcity of negative embodied data which is costly to collect and often filtered out or withheld in existing robotics datasets. Furthermore, we show that even modest exposure to real bad behavior data can improve alignment with human preferences and reduce costly false positives. We therefore call on the embodied AI community to curate and release their bad robot data, build synthetic bad data generation engines, develop more decentralized physical evaluation systems, and design benchmarks for fine-grained embodied reward model evaluations.

Prompt learning is a new machine learning paradigm that has attracted ample attention due to its simplicity and proven efficacy. Despite its growing adoption, the security vulnerabilities associated with this paradigm remain underexplored. In this work, we take the first step to propose BadBone, a stealthy and adaptive backdoor attack against prompt learning using bi-level optimization. Instead of backdooring the prompt learning process, we aim to compromise a backbone model such that only target downstream tasks employing prompt learning inherit the backdoor vulnerability. Extensive experiments on three different models and three datasets from various domains show that our targeted/untargeted backdoored models achieve high attack performance while maintaining utility on both pre-training and downstream tasks. Moreover, we evaluate our approach against six state-of-the-art model-level defenses, including Neural Cleanse, ABS, MNTD, NAD, CLP, and D-BR. The results demonstrate that these defenses are largely ineffective against our backdoored models and thus leave the effective defense as an important direction for future work.

Test-time scaling for vision-language-action (VLA) policies, methods such as RoboMonkey, SEAL, MG-Select, and V-GPS, samples K candidate action chunks at inference and executes the verifier-best. When all K candidates are unsafe, the system executes a violating action with no warning. We propose BOKBO, the first conformal abstention layer for K-sample VLA inference, providing finite-sample distribution-free guarantees on executed-violation rate. We provide both global and per-task (Mondrian) variants, with the per-task variant closing the conditional gap on the hardest tasks. Our analysis exposes a structural failure of policy-internal nonconformity scores under perturbation-based K-sampling: the base-policy confidence proxy and K-sample disagreement correlate at 0.98 with the action-noise hyperparameter $σ$, while correlating at the noise floor with actual safety violations. We test the failure's scope by replicating the analysis under token-level temperature sampling and find the failure is mechanism-specific and partially mitigated under policy-stochasticity-based sampling. A learned violation predictor conditioned on semantic visual features and task identity supports tight calibration: at $ε$ = 0.05 on libero_object_temp_x0.1 with OpenVLA-OFT, the conditional CRC bound holds on 86% of bootstrap splits with 78% coverage and 70% net task success. Mondrian-BOKBO raises the minimum per-task conditional hold fraction from 0.71 to 0.93. Results are stable across 5 training seeds, replicate within bootstrap noise on $π_0$-FAST, hold on libero_spatial_temp_x0.1 as a co-equal benchmark, and survive four within-suite distribution shifts. We additionally identify and correct a methodological pitfall: globally-set force thresholds well below expert-typical manipulation forces conflate unsafe behavior with normal manipulation, inflating violation rates by $5\times$.

Autonomous AI research agents aim to accelerate scientific discovery by automating the research pipeline, from hypothesis generation to peer review. However, existing benchmarks rarely test a fundamental bottleneck: whether Large Language Models can judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions, labeled with reviewer soundness sub-scores, and audited against source papers. SoundnessBench should be interpreted as a benchmark for recoverable proposal-stage soundness rather than exact prediction of full-paper review outcomes. Across 12 frontier LLMs, we find a pervasive optimism bias: under standard prompting, models frequently rate low-soundness proposals as sound, while aggressive prompting largely shifts errors from false positives to false negatives. Additional controls for public-corpus contamination, paper-identifying phrases, surface features, and human audit quality suggest that this behavior is not explained by a single confounder. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor.

Under standard graphical assumptions, the Markov boundary of a target variable is the smallest set of features that renders every other feature redundant. Once the boundary is observed, the target is conditionally independent of the rest of the table. This is a tempting object for tabular prediction, since it names exactly the columns a model should need. Yet modern regressors are still trained on the full feature set. We ask whether the Markov boundary is genuinely useful for prediction on SCM3K, a 3,450-task synthetic SCM benchmark with feature counts from 40 to 1000 and six SCM families, evaluated with six regressors. The answer is more nuanced than the theory suggests. Restricting a regressor to the oracle boundary often improves prediction substantially, and the improvement grows as the feature space becomes larger and sparser. But the natural pipeline of recovering the boundary with causal discovery and training on the recovered mask does not deliver. Existing estimators exhaust the compute budget before reaching the regime where the boundary helps most, and even where they run they rarely beat the full feature set. We trace this to three causes. Discovery optimizes structural recovery rather than prediction. False negatives and false positives carry sharply asymmetric predictive cost. The exact boundary is only one of many feature sets that beat all features. We then develop what these facts imply for prediction-aligned feature selection and for tabular models that learn to use causal structure.

Symbolic Regression (SR) plays a central role in scientific knowledge discovery by distilling mathematical equations from observational data. Most existing SR methods function within a bi-level optimization framework: an outer loop that searches for the discrete equation structure, and an inner loop that optimizes the continuous parameters of that structure. Crucially, parameter-fitting quality directly determines a structure's score and thus the outer-loop search. However, nonlinear operators make the inner loop highly non-convex, and budget-driven reliance on fast local solvers (e.g., BFGS) often yields poor local minima and underestimated scores for correct structures. This ``Good Structure, Bad Score'' phenomenon becomes a key bottleneck, degrading efficiency and misguiding the search away from the true equation. To resolve this, we propose SAGE-Fit (Structure-Aware and Semantics-Guided Evaluator for Symbolic Regression), an SR-native fitting framework that exploits the dual native priors of symbolic expressions. By capitalizing on the structural and semantic priors unique to SR, we design tailored modules for each property, thereby effectively mitigating this optimization bottleneck. Extensive experiments demonstrate that our approach, as a plug-and-play module, significantly enhances evaluation fidelity and universally improves the performance of various SR systems.

Achieving robust perception-reasoning synergy is a central goal for advanced Vision-Language Models (VLMs). Recent advancements have pursued this goal via architectural designs or agentic workflows. However, these approaches are often limited by static textual reasoning or complicated by the significant compute and engineering burden of external agentic complexity. Worse, this heavy investment does not yield proportional gains, often witnessing a "seesaw effect" on perception and reasoning. This motivates a fundamental rethinking of the true bottleneck. In this paper, we argue that the root cause of this trade-off is an ambiguity in modality credit assignment: when a VLM fails, is it due to flawed perception ("bad seeing") or flawed logic ("bad thinking")? To resolve this, we introduce a reinforcement learning framework that improves perception-reasoning synergy by reliably rewarding the perception fidelity. We explicitly decompose the generation process into interleaved perception and reasoning steps. This decoupling enables targeted supervision on perception. Crucially, we introduce Perception Verification (PV), leveraging a "blindfolded reasoning" proxy to reward perceptual fidelity independently of reasoning outcomes. Furthermore, to scale training across free-form VL tasks, we propose Structured Verbal Verification, which replaces high-variance LLM judging with structured algorithmic execution. These techniques are integrated into a Modality-Aware Credit Assignment (MoCA) mechanism, which routes rewards to the specific source of error -- either bad seeing or bad thinking -- enabling a single VLM to achieve simultaneous performance gains across a wide task spectrum.

Recent knowledge graph (KG)-enhanced large language models (LLMs) move beyond purely textual knowledge augmentation by encoding retrieved subgraphs into continuous soft prompts via graph neural networks, introducing a graph-conditioned channel that operates alongside the standard text interface. However, existing backdoor attacks are largely designed for the textual channel, and their effectiveness against this dual-channel architecture remains unclear. We show that this architecture creates a robustness gap: text-channel backdoor attacks that readily compromise textual KG prompting systems become largely ineffective against soft-prompt-based counterparts. We interpret this gap through semantic anchoring, whereby graph-derived soft prompts bias the generation-driving hidden state toward query-consistent semantics and suppress surface-level malicious instructions. Because this anchoring effect is itself induced by the graph channel, an attacker who manipulates graph-level representations can in turn redirect it toward adversarial semantics. To demonstrate this risk, we propose BadSKP, a backdoor attack that targets the graph-to-prompt interface through a multi-stage optimization strategy: it constructs adversarial target embeddings, optimizes poisoned node embeddings to steer the induced soft prompt, and approximates the optimized representations with fluent adversarial node attributes. Experiments on two soft-prompt KG-enhanced LLMs across four datasets show that BadSKP achieves high attack success under both frozen and trojaned settings, while text-only attacks remain unreliable even under perplexity-based defenses.

Imitation learning offers a promising framework for enabling robots to acquire diverse skills from human users. However, most imitation learning algorithms assume access to high-quality demonstrations an unrealistic expectation when collecting data from non-expert users, whose demonstrations often contain inadvertent errors. Naively learning from such demonstrations can result in unsafe policy behavior, while discarding entire demonstrations due to occasional mistakes wastes valuable data, especially in low-data settings. In this work, we introduce GiB (Good-in-Bad), an algorithm that automatically identifies and discards erroneous subtasks within demonstrations while preserving high-quality subtasks. The filtered data can then be used by any policy learning algorithm to train more robust policies. GiB first trains a self-supervised model to learn latent features and assigns binary weights to label each demonstration as good or bad. It then models the latent feature distribution of high-quality segments and uses the Mahalanobis distance to detect and evaluate poor-quality subtasks. We validate GiB on the Franka robot in both simulated and real-world multi-step tasks, demonstrating improved policy performance when learning from mixed-quality human demonstrations.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

The growing application of large language models (LLMs) in safety-critical domains has raised urgent concerns about their security. Many recent studies have demonstrated the feasibility of backdoor attacks against LLMs. However, existing methods suffer from three key shortcomings: explicit trigger patterns that compromise naturalness, unreliable injection of attacker-specified payloads in long-form generation, and incompletely specified threat models that obscure how backdoors are delivered and activated in practice. To address these gaps, we present BadStyle, a complete backdoor attack framework and pipeline. BadStyle leverages an LLM as a poisoned sample generator to construct natural and stealthy poisoned samples that carry imperceptible style-level triggers while preserving semantics and fluency. To stabilize payload injection during fine-tuning, we design an auxiliary target loss that reinforces the attacker-specified target content in responses to poisoned inputs and penalizes its emergence in benign responses. We further ground the attack in a realistic threat model and systematically evaluate BadStyle under both prompt-induced and PEFT-based injection strategies. Extensive experiments across seven victim LLMs, including LLaMA, Phi, DeepSeek, and GPT series, demonstrate that BadStyle achieves high attack success rates (ASRs) while maintaining strong stealthiness. The proposed auxiliary target loss substantially improves the stability of backdoor activation, yielding an average ASR improvement of around 30% across style-level triggers. Even in downstream deployment scenarios unknown during injection, the implanted backdoor remains effective. Moreover, BadStyle consistently evades representative input-level defenses and bypasses output-level defenses through simple camouflage.

Federated Learning (FL) systems are susceptible to adversarial attacks, such as model poisoning attacks and backdoor attacks. Existing defense mechanisms face critical limitations in deployments, such as relying on impractical assumptions (e.g., adversaries acknowledging the presence of attacks before attacking) or undermining accuracy in model training, even in benign scenarios. To address these challenges, we propose CustodianFL, a two-staged anomaly detection method specifically designed for FL deployments. In the first stage, it flags suspicious client activities. In the second stage that is activated only when needed, it further examines these candidates using Three-Sigma Rule to identify and exclude truly malicious local models from FL training. To ensure integrity and transparency within the FL system, CustodianFL integrates zero-knowledge proofs, enabling clients to cryptographically verify the server's detection process without relying on the server's goodwill. CustodianFL operates without unrealistic assumptions and avoids interfering with FL training in attack-free scenarios. It bridges the gap between theoretical advances in FL security and the practical demands of real FL systems. Experimental results demonstrate that CustodianFL consistently delivers performance comparable to benign cases, highlighting its effectiveness in identifying and eliminating malicious models with high accuracy.