As AI systems approach superhuman capabilities, scalable oversight increasingly relies on LLM-as-a-judge frameworks where models evaluate and guide each other's training. A core assumption is that binary preference labels provide only semantic supervision about response quality. We challenge this assumption by demonstrating that preference labels can function as a covert communication channel. We show that even when a neutral student model generates semantically unbiased completions, a biased judge can transmit unintended behavioral traits through preference assignments, which even strengthen across iterative alignment rounds. Our findings suggest that robust oversight in superalignment settings requires mechanisms that can detect and mitigate subliminal preference transmission, particularly when judges may pursue unintended objectives.

Jailbreak attacks remain a persistent threat to large language model safety. We propose Context-Conditioned Delta Steering (CC-Delta), an SAE-based defense that identifies jailbreak-relevant sparse features by comparing token-level representations of the same harmful request with and without jailbreak context. Using paired harmful/jailbreak prompts, CC-Delta selects features via statistical testing and applies inference-time mean-shift steering in SAE latent space. Across four aligned instruction-tuned models and twelve jailbreak attacks, CC-Delta achieves comparable or better safety-utility tradeoffs than baseline defenses operating in dense latent space. In particular, our method clearly outperforms dense mean-shift steering on all four models, and particularly against out-of-distribution attacks, showing that steering in sparse SAE feature space offers advantages over steering in dense activation space for jailbreak mitigation. Our results suggest off-the-shelf SAEs trained for interpretability can be repurposed as practical jailbreak defenses without task-specific training.

As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied data sets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To this end, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks and latent-space representation attacks; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. TamperBench requires minimal additional code to specify any fine-tuning configuration, alignment-stage defense method, and metric suite while ensuring end-to-end reproducibility. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. This yields novel insights, including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that Triplet emerges as a leading alignment-stage defense. Code is available at: https://github.com/criticalml-uw/TamperBench

LLM watermarks allow tracing AI-generated texts by inserting a detectable signal into their generated content. Recent works have proposed a wide range of watermarking algorithms, each with distinct designs, usually built using a bottom-up approach. Crucially, there is no general and principled formulation for LLM watermarking. In this work, we show that most existing and widely used watermarking schemes can in fact be derived from a principled constrained optimization problem. Our formulation unifies existing watermarking methods and explicitly reveals the constraints that each method optimizes. In particular, it highlights an understudied quality-diversity-power trade-off. At the same time, our framework also provides a principled approach for designing novel watermarking schemes tailored to specific requirements. For instance, it allows us to directly use perplexity as a proxy for quality, and derive new schemes that are optimal with respect to this constraint. Our experimental evaluation validates our framework: watermarking schemes derived from a given constraint consistently maximize detection power with respect to that constraint.

Prompt injection is one of the most critical vulnerabilities in LLM agents; yet, effective automated attacks remain largely unexplored from an optimization perspective. Existing methods heavily depend on human red-teamers and hand-crafted prompts, limiting their scalability and adaptability. We propose AutoInject, a reinforcement learning framework that generates universal, transferable adversarial suffixes while jointly optimizing for attack success and utility preservation on benign tasks. Our black-box method supports both query-based optimization and transfer attacks to unseen models and tasks. Using only a 1.5B parameter adversarial suffix generator, we successfully compromise frontier systems including GPT 5 Nano, Claude Sonnet 3.5, and Gemini 2.5 Flash on the AgentDojo benchmark, establishing a stronger baseline for automated prompt injection research.

Existing approaches for watermarking AI-generated images often rely on post-hoc methods applied in pixel space, introducing computational overhead and potential visual artifacts. In this work, we explore latent space watermarking and introduce DistSeal, a unified approach for latent watermarking that works across both diffusion and autoregressive models. Our approach works by training post-hoc watermarking models in the latent space of generative models. We demonstrate that these latent watermarkers can be effectively distilled either into the generative model itself or into the latent decoder, enabling in-model watermarking. The resulting latent watermarks achieve competitive robustness while offering similar imperceptibility and up to 20x speedup compared to pixel-space baselines. Our experiments further reveal that distilling latent watermarkers outperforms distilling pixel-space ones, providing a solution that is both more efficient and more robust.

Large Language Model (LLM)-based agent systems are increasingly deployed for complex real-world tasks but remain vulnerable to natural language-based attacks that exploit over-privileged tool use. This paper aims to understand and mitigate such attacks through the lens of privilege escalation, defined as agent actions exceeding the least privilege required for a user's intended task. Based on a formal model of LLM agent systems, we identify novel privilege escalation scenarios, particularly in multi-agent systems, including a variant akin to the classic confused deputy problem. To defend against both known and newly demonstrated privilege escalation, we propose SEAgent, a mandatory access control (MAC) framework built upon attribute-based access control (ABAC). SEAgent monitors agent-tool interactions via an information flow graph and enforces customizable security policies based on entity attributes. Our evaluations show that SEAgent effectively blocks various privilege escalation while maintaining a low false positive rate and negligible system overhead. This demonstrates its robustness and adaptability in securing LLM-based agent systems.

AI agents are vulnerable to prompt injection attacks, where malicious content hijacks agent behavior to steal credentials or cause financial loss. The only known robust defense is architectural isolation that strictly separates trusted task planning from untrusted environment observations. However, applying this design to Computer Use Agents (CUAs) -- systems that automate tasks by viewing screens and executing actions -- presents a fundamental challenge: current agents require continuous observation of UI state to determine each action, conflicting with the isolation required for security. We resolve this tension by demonstrating that UI workflows, while dynamic, are structurally predictable. We introduce Single-Shot Planning for CUAs, where a trusted planner generates a complete execution graph with conditional branches before any observation of potentially malicious content, providing provable control flow integrity guarantees against arbitrary instruction injections. Although this architectural isolation successfully prevents instruction injections, we show that additional measures are needed to prevent Branch Steering attacks, which manipulate UI elements to trigger unintended valid paths within the plan. We evaluate our design on OSWorld, and retain up to 57% of the performance of frontier models while improving performance for smaller open-source models by up to 19%, demonstrating that rigorous security and utility can coexist in CUAs.

Randomized Smoothing (RS) is a prominent technique for certifying the robustness of neural networks against adversarial perturbations. With RS, achieving high accuracy at small radii requires a small noise variance, while achieving high accuracy at large radii requires a large noise variance. However, the global noise variance used in the standard RS formulation leads to a fundamental limitation: there exists no global noise variance that simultaneously achieves strong performance at both small and large radii. To break through the global variance limitation, we propose a dual RS framework which enables input-dependent noise variances. To achieve that, we first prove that RS remains valid with input-dependent noise variances, provided the variance is locally constant around each input. Building on this result, we introduce two components which form our dual RS framework: (i) a variance estimator first predicts an optimal noise variance for each input, (ii) this estimated variance is then used by a standard RS classifier. The variance estimator is independently smoothed via RS to ensure local constancy, enabling flexible design. We also introduce training strategies to iteratively optimize the two components. Extensive experiments on CIFAR-10 show that our dual RS method provides strong performance for both small and large radii-unattainable with global noise variance-while incurring only a 60% computational overhead at inference. Moreover, it consistently outperforms prior input-dependent noise approaches across most radii, with particularly large gains at radii 0.5, 0.75, and 1.0, achieving relative improvements of 19%, 24%, and 21%, respectively. On ImageNet, dual RS remains effective across all radii. Additionally, the dual RS framework naturally provides a routing perspective for certified robustness, improving the accuracy-robustness trade-off with off-the-shelf expert RS models.

Federated Learning has seen an increased deployment in real-world scenarios recently, as it enables the distributed training of machine learning models without explicit data sharing between individual clients. Yet, the introduction of the so-called gradient inversion attacks has fundamentally challenged its privacy-preserving properties. Unfortunately, as these attacks mostly rely on direct data optimization without any formal guarantees, the vulnerability of real-world systems remains in dispute and requires tedious testing for each new federated deployment. To overcome these issues, recently the SPEAR attack was introduced, which is based on a theoretical analysis of the gradients of linear layers with ReLU activations. While SPEAR is an important theoretical breakthrough, the attack's practicality was severely limited by its exponential runtime in the batch size b. In this work, we fill this gap by applying State-of-the-Art techniques from Sparsely-Used Dictionary Learning to make the problem of gradient inversion on linear layers with ReLU activations tractable. Our experiments demonstrate that our new attack, SPEAR++, retains all desirable properties of SPEAR, such as robustness to DP noise and FedAvg aggregation, while being applicable to 10x bigger batch sizes.

Recently, semantically constrained adversarial examples (SemanticAE), which are directly generated from natural language instructions, have become a promising avenue for future research due to their flexible attacking forms. To generate SemanticAEs, current methods fall short of satisfactory attacking ability as the key underlying factors of semantic uncertainty in human instructions, such as referring diversity, descriptive incompleteness, and boundary ambiguity, have not been fully investigated. To tackle the issues, this paper develops a multi-dimensional instruction uncertainty reduction (InSUR) framework to generate more satisfactory SemanticAE, i.e., transferable, adaptive, and effective. Specifically, in the dimension of the sampling method, we propose the residual-driven attacking direction stabilization to alleviate the unstable adversarial optimization caused by the diversity of language references. By coarsely predicting the language-guided sampling process, the optimization process will be stabilized by the designed ResAdv-DDIM sampler, therefore releasing the transferable and robust adversarial capability of multi-step diffusion models. In task modeling, we propose the context-encoded attacking scenario constraint to supplement the missing knowledge from incomplete human instructions. Guidance masking and renderer integration are proposed to regulate the constraints of 2D/3D SemanticAE, activating stronger scenario-adapted attacks. Moreover, in the dimension of generator evaluation, we propose the semantic-abstracted attacking evaluation enhancement by clarifying the evaluation boundary, facilitating the development of more effective SemanticAE generators. Extensive experiments demonstrate the superiority of the transfer attack performance of InSUR. Moreover, we realize the reference-free generation of semantically constrained 3D adversarial examples for the first time.

AI agents powered by large language models (LLMs) are being deployed at scale, yet we lack a systematic understanding of how the choice of backbone LLM affects agent security. The non-deterministic sequential nature of AI agents complicates security modeling, while the integration of traditional software with AI components entangles novel LLM vulnerabilities with conventional security risks. Existing frameworks only partially address these challenges as they either capture specific vulnerabilities only or require modeling of complete agents. To address these limitations, we introduce threat snapshots: a framework that isolates specific states in an agent's execution flow where LLM vulnerabilities manifest, enabling the systematic identification and categorization of security risks that propagate from the LLM to the agent level. We apply this framework to construct the $\operatorname{b}^3$ benchmark, a security benchmark based on 194331 unique crowdsourced adversarial attacks. We then evaluate 31 popular LLMs with it, revealing, among other insights, that enhanced reasoning capabilities improve security, while model size does not correlate with security. We release our benchmark, dataset, and evaluation code to facilitate widespread adoption by LLM providers and practitioners, offering guidance for agent developers and incentivizing model developers to prioritize backbone security improvements.

Recent years have seen a surge in interest in digital content watermarking techniques, driven by the proliferation of generative models and increased legal pressure. With an ever-growing percentage of AI-generated content available online, watermarking plays an increasingly important role in ensuring content authenticity and attribution at scale. There have been many works assessing the robustness of watermarking to removal attacks, yet, watermark forging, the scenario when a watermark is stolen from genuine content and applied to malicious content, remains underexplored. In this work, we investigate watermark forging in the context of widely used post-hoc image watermarking. Our contributions are as follows. First, we introduce a preference model to assess whether an image is watermarked. The model is trained using a ranking loss on purely procedurally generated images without any need for real watermarks. Second, we demonstrate the model's capability to remove and forge watermarks by optimizing the input image through backpropagation. This technique requires only a single watermarked image and works without knowledge of the watermarking model, making our attack much simpler and more practical than attacks introduced in related work. Third, we evaluate our proposed method on a variety of post-hoc image watermarking models, demonstrating that our approach can effectively forge watermarks, questioning the security of current watermarking approaches. Our code and further resources are publicly available.

Recent findings suggest that misaligned models may exhibit deceptive behavior, raising concerns about output trustworthiness. Chain-of-thought (CoT) is a promising tool for alignment monitoring: when models articulate their reasoning faithfully, monitors can detect and mitigate harmful behaviors before undesirable outcomes occur. However, a key uncertainty is: Can models obfuscate their CoT in order to pursue hidden adversarial objectives while evading detection? To answer this question and thus stress-test CoT monitorability, we develop a composable and quantifiable taxonomy of prompts to elicit CoT obfuscation. We evaluate both internal CoT (reasoning traces) and external CoT (prompted reasoning in outputs) using toy tasks and more realistic environments in SHADE-Arena. We show that: (i) CoT monitoring performs accurately and efficiently without obfuscation pressure. (ii) Under strong obfuscation pressure, some models successfully complete adversarial tasks while evading detection. (iii) Models do not obfuscate their internal CoT as much as their external CoT (under prompt pressure). These results suggest that while CoT provides valuable oversight in benign settings, robust deployment requires model-specific stress-testing of monitorability.

LLMs are often used by downstream users as teacher models for knowledge distillation, compressing their capabilities into memory-efficient models. However, as these teacher models may stem from untrusted parties, distillation can raise unexpected security risks. In this paper, we investigate the security implications of knowledge distillation from backdoored teacher models. First, we show that prior backdoors mostly do not transfer onto student models. Our key insight is that this is because existing LLM backdooring methods choose trigger tokens that rarely occur in usual contexts. We argue that this underestimates the security risks of knowledge distillation and introduce a new backdooring technique, T-MTB, that enables the construction and study of transferable backdoors. T-MTB carefully constructs a composite backdoor trigger, made up of several specific tokens that often occur individually in anticipated distillation datasets. As such, the poisoned teacher remains stealthy, while during distillation the individual presence of these tokens provides enough signal for the backdoor to transfer onto the student. Using T-MTB, we demonstrate and extensively study the security risks of transferable backdoors across two attack scenarios, jailbreaking and content modulation, and across four model families of LLMs.

We present a novel approach for attacking black-box large language models (LLMs) by exploiting their ability to express confidence in natural language. Existing black-box attacks require either access to continuous model outputs like logits or confidence scores (which are rarely available in practice), or rely on proxy signals from other models. Instead, we demonstrate how to prompt LLMs to express their internal confidence in a way that is sufficiently calibrated to enable effective adversarial optimization. We apply our general method to three attack scenarios: adversarial examples for vision-LLMs, jailbreaks and prompt injections. Our attacks successfully generate malicious inputs against systems that only expose textual outputs, thereby dramatically expanding the attack surface for deployed LLMs. We further find that better and larger models exhibit superior calibration when expressing confidence, creating a concerning security paradox where model capability improvements directly enhance vulnerability. Our code is available at this [link](https://github.com/zj-jayzhang/black_box_llm_optimization).

Text-to-image (T2I) models, though exhibiting remarkable creativity in image generation, can be exploited to produce unsafe images. Existing safety measures, e.g., content moderation or model alignment, fail in the presence of white-box adversaries who know and can adjust model parameters, e.g., by fine-tuning. This paper presents a novel defensive framework, named Patronus, which equips T2I models with holistic protection to defend against white-box adversaries. Specifically, we design an internal moderator that decodes unsafe input features into zero vectors while ensuring the decoding performance of benign input features. Furthermore, we strengthen the model alignment with a carefully designed non-fine-tunable learning mechanism, ensuring the T2I model will not be compromised by malicious fine-tuning. We conduct extensive experiments to validate the intactness of the performance on safe content generation and the effectiveness of rejecting unsafe content generation. Results also confirm the resilience of Patronus against various fine-tuning attacks by white-box adversaries.

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed. Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques-gradient descent, reinforcement learning, random search, and human-guided exploration-we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

The widespread application of large vision language models has significantly raised safety concerns. In this project, we investigate text prompt injection, a simple yet effective method to mislead these models. We developed an algorithm for this type of attack and demonstrated its effectiveness and efficiency through experiments. Compared to other attack methods, our approach is particularly effective for large models without high demand for computational resources.

The increasing adoption of LLM agents with access to numerous tools and sensitive data significantly widens the attack surface for indirect prompt injections. Due to the context-dependent nature of attacks, however, current defenses are often ill-calibrated as they cannot reliably differentiate malicious and benign instructions, leading to high false positive rates that prevent their real-world adoption. To address this, we present a novel approach inspired by the fundamental principle of computer security: data should not contain executable instructions. Instead of sample-level classification, we propose a token-level sanitization process, which surgically removes any instructions directed at AI systems from tool outputs, capturing malicious instructions as a byproduct. In contrast to existing safety classifiers, this approach is non-blocking, does not require calibration, and is agnostic to the context of tool outputs. Further, we can train such token-level predictors with readily available instruction-tuning data only, and don't have to rely on unrealistic prompt injection examples from challenges or of other synthetic origin. In our experiments, we find that this approach generalizes well across a wide range of attacks and benchmarks like AgentDojo, BIPIA, InjecAgent, ASB and SEP, achieving a 7-10x reduction of attack success rate (ASR) (34% to 3% on AgentDojo), without impairing agent utility in both benign and malicious settings.