Fine-tuning APIs offered by major AI providers create new attack surfaces where adversaries can bypass safety measures through targeted fine-tuning. We introduce Trojan-Speak, an adversarial fine-tuning method that bypasses Anthropic's Constitutional Classifiers. Our approach uses curriculum learning combined with GRPO-based hybrid reinforcement learning to teach models a communication protocol that evades LLM-based content classification. Crucially, while prior adversarial fine-tuning approaches report more than 25% capability degradation on reasoning benchmarks, Trojan-Speak incurs less than 5% degradation while achieving 99+% classifier evasion for models with 14B+ parameters. We demonstrate that fine-tuned models can provide detailed responses to expert-level CBRN (Chemical, Biological, Radiological, and Nuclear) queries from Anthropic's Constitutional Classifiers bug-bounty program. Our findings reveal that LLM-based content classifiers alone are insufficient for preventing dangerous information disclosure when adversaries have fine-tuning access, and we show that activation-level probes can substantially improve robustness to such attacks.

Test-time training (TTT) has recently emerged as a promising method to improve the reasoning abilities of large language models (LLMs), in which the model directly learns from test data without access to labels. However, this reliance on test data also makes TTT methods vulnerable to harmful prompt injections. In this paper, we investigate safety vulnerabilities of TTT methods, where we study a representative self-consistency-based test-time learning method: test-time reinforcement learning (TTRL), a recent TTT method that improves LLM reasoning by rewarding self-consistency using majority vote as a reward signal. We show that harmful prompt injection during TTRL amplifies the model's existing behaviors, i.e., safety amplification when the base model is relatively safe, and harmfulness amplification when it is vulnerable to the injected data. In both cases, there is a decline in reasoning ability, which we refer to as the reasoning tax. We also show that TTT methods such as TTRL can be exploited adversarially using specially designed "HarmInject" prompts to force the model to answer jailbreak and reasoning queries together, resulting in stronger harmfulness amplification. Overall, our results highlight that TTT methods that enhance LLM reasoning by promoting self-consistency can lead to amplification behaviors and reasoning degradation, highlighting the need for safer TTT methods.

Cloud-based infrastructures have become the dominant platform for deploying large models, particularly large language models (LLMs). Fine-tuning and inference are increasingly delegated to cloud providers for simplified deployment and access to proprietary models, yet this creates a fundamental trust gap: although cryptographic and TEE-based verification exist, the scale of modern LLMs renders them prohibitive, leaving clients unable to practically audit these processes. This lack of transparency creates concrete security risks that can silently compromise service integrity. We present AFTUNE, an auditable and verifiable framework that ensures the computation integrity of cloud-based fine-tuning and inference. AFTUNE incorporates a lightweight recording and spot-check mechanism that produces verifiable traces of execution. These traces enable clients to later audit whether the training and inference processes followed the agreed configurations. Our evaluation shows that AFTUNE imposes practical computation overhead while enabling selective and efficient verification, demonstrating that trustworthy model services are achievable in today's cloud environments.

Current safety mechanisms for Large Language Models (LLMs) rely heavily on static, fine-tuned classifiers that suffer from adaptation rigidity, the inability to enforce new governance rules without expensive retraining. To address this, we introduce CourtGuard, a retrieval-augmented multi-agent framework that reimagines safety evaluation as Evidentiary Debate. By orchestrating an adversarial debate grounded in external policy documents, CourtGuard achieves state-of-the-art performance across 7 safety benchmarks, outperforming dedicated policy-following baselines without fine-tuning. Beyond standard metrics, we highlight two critical capabilities: (1) Zero-Shot Adaptability, where our framework successfully generalized to an out-of-domain Wikipedia Vandalism task (achieving 90\% accuracy) by swapping the reference policy; and (2) Automated Data Curation and Auditing, where we leveraged CourtGuard to curate and audit nine novel datasets of sophisticated adversarial attacks. Our results demonstrate that decoupling safety logic from model weights offers a robust, interpretable, and adaptable path for meeting current and future regulatory requirements in AI governance.

Advances in Generative AI (GenAI) have led to the development of various protection strategies to prevent the unauthorized use of images. These methods rely on adding imperceptible protective perturbations to images to thwart misuse such as style mimicry or deepfake manipulations. Although previous attacks on these protections required specialized, purpose-built methods, we demonstrate that this is no longer necessary. We show that off-the-shelf image-to-image GenAI models can be repurposed as generic ``denoisers" using a simple text prompt, effectively removing a wide range of protective perturbations. Across 8 case studies spanning 6 diverse protection schemes, our general-purpose attack not only circumvents these defenses but also outperforms existing specialized attacks while preserving the image's utility for the adversary. Our findings reveal a critical and widespread vulnerability in the current landscape of image protection, indicating that many schemes provide a false sense of security. We stress the urgent need to develop robust defenses and establish that any future protection mechanism must be benchmarked against attacks from off-the-shelf GenAI models. Code is available in this repository: https://github.com/mlsecviswanath/img2imgdenoiser

Trusted Execution Environments (TEEs) (e.g., Intel SGX and ArmTrustZone) aim to protect sensitive computation from a compromised operating system, yet real deployments remain vulnerable to microarchitectural leakage, side-channel attacks, and fault injection. In parallel, security teams increasingly rely on Large Language Model (LLM) assistants as security advisors for TEE architecture review, mitigation planning, and vulnerability triage. This creates a socio-technical risk surface: assistants may hallucinate TEE mechanisms, overclaim guarantees (e.g., what attestation does and does not establish), or behave unsafely under adversarial prompting. We present a red-teaming study of two prevalently deployed LLM assistants in the role of TEE security advisors: ChatGPT-5.2 and Claude Opus-4.6, focusing on the inherent limitations and transferability of prompt-induced failures across LLMs. We introduce TEE-RedBench, a TEE-grounded evaluation methodology comprising (i) a TEE-specific threat model for LLM-mediated security work, (ii) a structured prompt suite spanning SGX and TrustZone architecture, attestation and key management, threat modeling, and non-operational mitigation guidance, along with policy-bound misuse probes, and (iii) an annotation rubric that jointly measures technical correctness, groundedness, uncertainty calibration, refusal quality, and safe helpfulness. We find that some failures are not purely idiosyncratic, transferring up to 12.02% across LLM assistants, and we connect these outcomes to secure architecture by outlining an "LLM-in-the-loop" evaluation pipeline: policy gating, retrieval grounding, structured templates, and lightweight verification checks that, when combined, reduce failures by 80.62%.

Addressing the critical need for robust safety in Large Language Models (LLMs), particularly against adversarial attacks and in-distribution errors, we introduce Reinforcement Learning with Backtracking Feedback (RLBF). This framework advances upon prior methods, such as BSAFE, by primarily leveraging a Reinforcement Learning (RL) stage where models learn to dynamically correct their own generation errors. Through RL with critic feedback on the model's live outputs, LLMs are trained to identify and recover from their actual, emergent safety violations by emitting an efficient "backtrack by x tokens" signal, then continuing generation autoregressively. This RL process is crucial for instilling resilience against sophisticated adversarial strategies, including middle filling, Greedy Coordinate Gradient (GCG) attacks, and decoding parameter manipulations. To further support the acquisition of this backtracking capability, we also propose an enhanced Supervised Fine-Tuning (SFT) data generation strategy (BSAFE+). This method improves upon previous data creation techniques by injecting violations into coherent, originally safe text, providing more effective initial training for the backtracking mechanism. Comprehensive empirical evaluations demonstrate that RLBF significantly reduces attack success rates across diverse benchmarks and model scales, achieving superior safety outcomes while critically preserving foundational model utility.

As AI agents automate critical workloads, they remain vulnerable to indirect prompt injection (IPI) attacks. Current defenses rely on monitoring protocols that jointly evaluate an agent's Chain-of-Thought (CoT) and tool-use actions to ensure alignment with user intent. We demonstrate that these monitoring-based defenses can be bypassed via a novel Agent-as-a-Proxy attack, where prompt injection attacks treat the agent as a delivery mechanism, bypassing both agent and monitor simultaneously. While prior work on scalable oversight has focused on whether small monitors can supervise large agents, we show that even frontier-scale monitors are vulnerable. Large-scale monitoring models like Qwen2.5-72B can be bypassed by agents with similar capabilities, such as GPT-4o mini and Llama-3.1-70B. On the AgentDojo benchmark, we achieve a high attack success rate against AlignmentCheck and Extract-and-Evaluate monitors under diverse monitoring LLMs. Our findings suggest current monitoring-based agentic defenses are fundamentally fragile regardless of model scale.

The rise of bot accounts on social media poses significant risks to public discourse. To address this threat, modern bot detectors increasingly rely on Graph Neural Networks (GNNs). However, the effectiveness of these GNN-based detectors in real-world settings remains poorly understood. In practice, attackers continuously adapt their strategies as well as must operate under domain-specific and temporal constraints, which can fundamentally limit the applicability of existing attack methods. As a result, there is a critical need for robust GNN-based bot detection methods under realistic, constraint-aware attack scenarios. To address this gap, we introduce BOCLOAK to systematically evaluate the robustness of GNN-based social bot detection via both edge editing and node injection adversarial attacks under realistic constraints. BOCLOAK constructs a probability measure over spatio-temporal neighbor features and learns an optimal transport geometry that separates human and bot behaviors. It then decodes transport plans into sparse, plausible edge edits that evade detection while obeying real-world constraints. We evaluate BOCLOAK across three social bot datasets, five state-of-the-art bot detectors, three adversarial defenses, and compare it against four leading graph adversarial attack baselines. BOCLOAK achieves up to 80.13% higher attack success rates while using 99.80% less GPU memory under realistic real-world constraints. Most importantly, BOCLOAK shows that optimal transport provides a lightweight, principled framework for bridging the gap between adversarial attacks and real-world bot detection.

Multimodal Large Language Models (MLLMs) have achieved remarkable performance across vision-language tasks. Recent advancements allow these models to process multiple images as inputs. However, the vulnerabilities of multi-image MLLMs remain unexplored. Existing adversarial attacks focus on single-image settings and often assume a white-box threat model, which is impractical in many real-world scenarios. This paper introduces LAMP, a black-box method for learning Universal Adversarial Perturbations (UAPs) targeting multi-image MLLMs. LAMP applies an attention-based constraint that prevents the model from effectively aggregating information across images. LAMP also introduces a novel cross-image contagious constraint that forces perturbed tokens to influence clean tokens, spreading adversarial effects without requiring all inputs to be modified. Additionally, an index-attention suppression loss enables a robust position-invariant attack. Experimental results show that LAMP outperforms SOTA baselines and achieves the highest attack success rates across multiple vision-language tasks and models.

Multimodal foundation models that integrate audio, vision, and language achieve strong performance on reasoning and generation tasks, yet their robustness to adversarial manipulation remains poorly understood. We study a realistic and underexplored threat model: untargeted, audio-only adversarial attacks on trimodal audio-video-language models. We analyze six complementary attack objectives that target different stages of multimodal processing, including audio encoder representations, cross-modal attention, hidden states, and output likelihoods. Across three state-of-the-art models and multiple benchmarks, we show that audio-only perturbations can induce severe multimodal failures, achieving up to 96% attack success rate. We further show that attacks can be successful at low perceptual distortions (LPIPS <= 0.08, SI-SNR >= 0) and benefit more from extended optimization than increased data scale. Transferability across models and encoders remains limited, while speech recognition systems such as Whisper primarily respond to perturbation magnitude, achieving >97% attack success under severe distortion. These results expose a previously overlooked single-modality attack surface in multimodal systems and motivate defenses that enforce cross-modal consistency.

Fine-tuning is an essential and pervasive functionality for applying large language models (LLMs) to downstream tasks. However, it has the potential to substantially degrade safety alignment, e.g., by greatly increasing susceptibility to jailbreak attacks, even when the fine-tuning data is entirely harmless. Despite garnering growing attention in defense efforts during the fine-tuning stage, existing methods struggle with a persistent safety-utility dilemma: emphasizing safety compromises task performance, whereas prioritizing utility typically requires deep fine-tuning that inevitably leads to steep safety declination. In this work, we address this dilemma by shedding new light on the geometric interaction between safety- and utility-oriented gradients in safety-aligned LLMs. Through systematic empirical analysis, we uncover three key insights: (I) safety gradients lie in a low-rank subspace, while utility gradients span a broader high-dimensional space; (II) these subspaces are often negatively correlated, causing directional conflicts during fine-tuning; and (III) the dominant safety direction can be efficiently estimated from a single sample. Building upon these novel insights, we propose safety-preserving fine-tuning (SPF), a lightweight approach that explicitly removes gradient components conflicting with the low-rank safety subspace. Theoretically, we show that SPF guarantees utility convergence while bounding safety drift. Empirically, SPF consistently maintains downstream task performance and recovers nearly all pre-trained safety alignment, even under adversarial fine-tuning scenarios. Furthermore, SPF exhibits robust resistance to both deep fine-tuning and dynamic jailbreak attacks. Together, our findings provide new mechanistic understanding and practical guidance toward always-aligned LLM fine-tuning.

Fine-tuning safety-aligned large language models (LLMs) can substantially compromise their safety. Previous approaches require many safety samples or calibration sets, which not only incur significant computational overhead during realignment but also lead to noticeable degradation in model utility. Contrary to this belief, we show that safety alignment can be fully recovered with only a single safety example, without sacrificing utility and at minimal cost. Remarkably, this recovery is effective regardless of the number of harmful examples used in fine-tuning or the size of the underlying model, and convergence is achieved within just a few epochs. Furthermore, we uncover the low-rank structure of the safety gradient, which explains why such efficient correction is possible. We validate our findings across five safety-aligned LLMs and multiple datasets, demonstrating the generality of our approach.

Large Language Models (LLMs) deployed in practical and safety-critical settings are increasingly susceptible to bit-flip faults caused by hardware degradation, cosmic radiation, or deliberate fault-injection attacks such as Rowhammer. These faults silently corrupt internal parameters and can lead to unpredictable or dangerous model behavior. Localizing these corruptions is essential: without identifying the affected region, it is impossible to diagnose the source of degradation, apply targeted corrective measures, or restore model functionality without resorting to costly fine-tuning or full retraining. This work introduces BitFlipScope, a scalable, software-based framework for identifying fault-affected regions within transformer architectures under two deployment scenarios. When a clean reference model is available, BitFlipScope performs differential analysis of outputs, hidden states, and internal activations for detecting anomalous behavior indicative of corruption to pinpoint or localize faults. When no reference model exists, it uses residual-path perturbation and loss-sensitivity profiling to infer the fault-impacted region directly from the corrupted model. In both settings, the framework not only enables effective fault diagnosis but also supports lightweight performance recovery without fine-tuning, offering a practical path to restoring corrupted models. Together, these capabilities make BitFlipScope an important step toward trustworthy, fault-resilient LLM deployment in hardware-prone and adversarial environments.

Federated Learning (FL) has emerged as a leading privacy-preserving machine learning paradigm, enabling participants to share model updates instead of raw data. However, FL continues to face key challenges, including weak client incentives, privacy risks, and resource constraints. Assessing client reliability is essential for fair incentive allocation and ensuring that each client's data contributes meaningfully to the global model. To this end, we propose MURIM, a MUlti-dimensional Reputation-based Incentive Mechanism that jointly considers client reliability, privacy, resource capacity, and fairness while preventing malicious or unreliable clients from earning undeserved rewards. MURIM allocates incentives based on client contribution, latency, and reputation, supported by a reliability verification module. Extensive experiments on MNIST, FMNIST, and ADULT Income datasets demonstrate that MURIM achieves up to 18% improvement in fairness metrics, reduces privacy attack success rates by 5-9%, and improves robustness against poisoning and noisy-gradient attacks by up to 85% compared to state-of-the-art baselines. Overall, MURIM effectively mitigates adversarial threats, promotes fair and truthful participation, and preserves stable model convergence across heterogeneous and dynamic federated settings.

Vertical Federated Learning (VFL) enables collaborative model training across organizations that share common user samples but hold disjoint feature spaces. Despite its potential, VFL is susceptible to feature inference attacks, in which adversarial parties exploit shared confidence scores (i.e., prediction probabilities) during inference to reconstruct private input features of other participants. To counter this threat, we propose PRIVEE (PRIvacy-preserving Vertical fEderated lEarning), a novel defense mechanism named after the French word privée, meaning "private." PRIVEE obfuscates confidence scores while preserving critical properties such as relative ranking and inter-score distances. Rather than exposing raw scores, PRIVEE shares only the transformed representations, mitigating the risk of reconstruction attacks without degrading model prediction accuracy. Extensive experiments show that PRIVEE achieves a threefold improvement in privacy protection compared to state-of-the-art defenses, while preserving full predictive performance against advanced feature inference attacks.

Multi-Object Tracking (MOT) is a critical task in computer vision, with applications ranging from surveillance systems to autonomous driving. However, threats to MOT algorithms have yet been widely studied. In particular, incorrect association between the tracked objects and their assigned IDs can lead to severe consequences, such as wrong trajectory predictions. Previous attacks against MOT either focused on hijacking the trackers of individual objects, or manipulating the tracker IDs in MOT by attacking the integrated object detection (OD) module in the digital domain, which are model-specific, non-robust, and only able to affect specific samples in offline datasets. In this paper, we present AdvTraj, the first online and physical ID-manipulation attack against tracking-by-detection MOT, in which an attacker uses adversarial trajectories to transfer its ID to a targeted object to confuse the tracking system, without attacking OD. Our simulation results in CARLA show that AdvTraj can fool ID assignments with 100% success rate in various scenarios for white-box attacks against SORT, which also have high attack transferability (up to 93% attack success rate) against state-of-the-art (SOTA) MOT algorithms due to their common design principles. We characterize the patterns of trajectories generated by AdvTraj and propose two universal adversarial maneuvers that can be performed by a human walker/driver in daily scenarios. Our work reveals under-explored weaknesses in the object association phase of SOTA MOT systems, and provides insights into enhancing the robustness of such systems.

The memorization of training data in large language models (LLMs) poses significant privacy and copyright concerns. Existing data extraction methods, particularly heuristic-based divergence attacks, often exhibit limited success and offer limited insight into the fundamental drivers of memorization leakage. This paper introduces Confusion-Inducing Attacks (CIA), a principled framework for extracting memorized data by systematically maximizing model uncertainty. We empirically demonstrate that the emission of memorized text during divergence is preceded by a sustained spike in token-level prediction entropy. CIA leverages this insight by optimizing input snippets to deliberately induce this consecutive high-entropy state. For aligned LLMs, we further propose Mismatched Supervised Fine-tuning (SFT) to simultaneously weaken their alignment and induce targeted confusion, thereby increasing susceptibility to our attacks. Experiments on various unaligned and aligned LLMs demonstrate that our proposed attacks outperform existing baselines in extracting verbatim and near-verbatim training data without requiring prior knowledge of the training data. Our findings highlight persistent memorization risks across various LLMs and offer a more systematic method for assessing these vulnerabilities.

Large language models remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Defending against novel jailbreaks represents a critical challenge in AI safety. Adversarial training -- designed to make models robust against worst-case perturbations -- has been the dominant paradigm for adversarial robustness. However, due to optimization challenges and difficulties in defining realistic threat models, adversarial training methods often fail on newly developed jailbreaks in practice. This paper proposes a new paradigm for improving robustness against unseen jailbreaks, centered on the Adversarial Déjà Vu hypothesis: novel jailbreaks are not fundamentally new, but largely recombinations of adversarial skills from previous attacks. We study this hypothesis through a large-scale analysis of 32 attack papers published over two years. Using an automated pipeline, we extract and compress adversarial skills into a sparse dictionary of primitives, with LLMs generating human-readable descriptions. Our analysis reveals that unseen attacks can be effectively explained as sparse compositions of earlier skills, with explanatory power increasing monotonically as skill coverage grows. Guided by this insight, we introduce Adversarial Skill Compositional Training (ASCoT), which trains on diverse compositions of skill primitives rather than isolated attack instances. ASCoT substantially improves robustness to unseen attacks, including multi-turn jailbreaks, while maintaining low over-refusal rates. We also demonstrate that expanding adversarial skill coverage, not just data scale, is key to defending against novel attacks. \textcolor{red}{\textbf{Warning: This paper contains content that may be harmful or offensive in nature.

Federated Learning (FL) has gained significant attention for its privacy-preserving capabilities, enabling distributed devices to collaboratively train a global model without sharing raw data. However, its distributed nature forces the central server to blindly trust the local training process and aggregate uncertain model updates, making it susceptible to Byzantine attacks from malicious participants, especially in mission-critical scenarios. Detecting such attacks is challenging due to the diverse knowledge across clients, where variations in model updates may stem from benign factors, such as non-IID data, rather than adversarial behavior. Existing data-driven defenses struggle to distinguish malicious updates from natural variations, leading to high false positive rates and poor filtering performance. To address this challenge, we propose Sentinel, a remote attestation (RA)-based scheme for FL systems that regains client-side transparency and mitigates Byzantine attacks from a system security perspective. Our system employs code instrumentation to track control-flow and monitor critical variables in the local training process. Additionally, we utilize a trusted training recorder within a Trusted Execution Environment (TEE) to generate an attestation report, which is cryptographically signed and securely transmitted to the server. Upon verification, the server ensures that legitimate client training processes remain free from program behavior violation or data manipulation, allowing only trusted model updates to be aggregated into the global model. Experimental results on IoT devices demonstrate that Sentinel ensures the trustworthiness of the local training integrity with low runtime and memory overhead.