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.

Contrastive steering has been shown as a simple and effective method to adjust the generative behavior of LLMs at inference time. It uses examples of prompt responses with and without a trait to identify a direction in an intermediate activation layer, and then shifts activations in this 1-dimensional subspace. However, despite its growing use in AI safety applications, the robustness of contrastive steering to noisy or adversarial data corruption is poorly understood. We initiate a study of the robustness of this process with respect to corruption of the dataset of examples used to train the steering direction. Our first observation is that contrastive steering is quite robust to a moderate amount of corruption, but unwanted side effects can be clearly and maliciously manifested when a non-trivial fraction of the training data is altered. Second, we analyze the geometry of various types of corruption, and identify some safeguards. Notably, a key step in learning the steering direction involves high-dimensional mean computation, and we show that replacing this step with a recently developed robust mean estimator often mitigates most of the unwanted effects of malicious corruption.

Alignment audits aim to robustly identify hidden goals from strategic, situationally aware misaligned models. Despite this threat model, existing auditing methods have not been systematically stress-tested against deception strategies. We address this gap, implementing an automatic red-team pipeline that generates deception strategies (in the form of system prompts) tailored to specific white-box and black-box auditing methods. Stress-testing assistant prefills, user persona sampling, sparse autoencoders, and token embedding similarity methods against secret-keeping model organisms, our automatic red-team pipeline finds prompts that deceive both the black-box and white-box methods into confident, incorrect guesses. Our results provide the first documented evidence of activation-based strategic deception, and suggest that current black-box and white-box methods would not be robust to a sufficiently capable misaligned model.

Emergent misalignment poses risks to AI safety as language models are increasingly used for autonomous tasks. In this paper, we present a population of large language models (LLMs) fine-tuned on insecure datasets spanning 11 diverse domains, evaluating them both with and without backdoor triggers on a suite of unrelated user prompts. Our evaluation experiments on \texttt{Qwen2.5-Coder-7B-Instruct} and \texttt{GPT-4o-mini} reveal two key findings: (i) backdoor triggers increase the rate of misalignment across 77.8% of domains (average drop: 4.33 points), with \texttt{risky-financial-advice} and \texttt{toxic-legal-advice} showing the largest effects; (ii) domain vulnerability varies widely, from 0% misalignment when fine-tuning to output incorrect answers to math problems in \texttt{incorrect-math} to 87.67% when fine-tuned on \texttt{gore-movie-trivia}. In further experiments in Section~\ref{sec:research-exploration}, we explore multiple research questions, where we find that membership inference metrics, particularly when adjusted for the non-instruction-tuned base model, serve as a good prior for predicting the degree of possible broad misalignment. Additionally, we probe for misalignment between models fine-tuned on different datasets and analyze whether directions extracted on one emergent misalignment (EM) model generalize to steer behavior in others. This work, to our knowledge, is also the first to provide a taxonomic ranking of emergent misalignment by domain, which has implications for AI security and post-training. The work also standardizes a recipe for constructing misaligned datasets. All code and datasets are publicly available on GitHub.\footnote{https://github.com/abhishek9909/assessing-domain-emergent-misalignment/tree/main}

Personalization is crucial for aligning Large Language Model (LLM) outputs with individual user preferences and background knowledge. State-of-the-art solutions are based on retrieval augmentation, where relevant context from a user profile is retrieved for LLM consumption. These methods deal with a trade-off between exposing retrieved private data to cloud providers and relying on less capable local models. We introduce $P^3$, an interactive framework for high-quality personalization without revealing private profiles to server-side LLMs. In $P^3$, a large server-side model generates a sequence of $k$ draft tokens based solely on the user query, while a small client-side model, with retrieval access to the user's private profile, evaluates and modifies these drafts to better reflect user preferences. This process repeats until an end token is generated. Experiments on LaMP-QA, a recent benchmark consisting of three personalized question answering datasets, show that $P^3$ consistently outperforms both non-personalized server-side and personalized client-side baselines, achieving statistically significant improvements of $7.4%$ to $9%$ on average. Importantly, $P^3$ recovers $90.3%$ to $95.7%$ of the utility of a ``leaky'' upper-bound scenario in which the full profile is exposed to the large server-side model. Privacy analyses, including linkability and attribute inference attacks, indicate that $P^3$ preserves the privacy of a non-personalized server-side model, introducing only marginal additional leakage ($1.5%$--$3.5%$) compared to submitting a query without any personal context. Additionally, the framework is efficient for edge deployment, with the client-side model generating only $9.2%$ of the total tokens. These results demonstrate that $P^3$ provides a practical, effective solution for personalized generation with improved privacy.

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.

Emoticons are widely used in digital communication to convey affective intent, yet their safety implications for Large Language Models (LLMs) remain largely unexplored. In this paper, we identify emoticon semantic confusion, a vulnerability where LLMs misinterpret ASCII-based emoticons to perform unintended and even destructive actions. To systematically study this phenomenon, we develop an automated data generation pipeline and construct a dataset containing 3,757 code-oriented test cases spanning 21 meta-scenarios, four programming languages, and varying contextual complexities. Our study on six LLMs reveals that emoticon semantic confusion is pervasive, with an average confusion ratio exceeding 38%. More critically, over 90% of confused responses yield 'silent failures', which are syntactically valid outputs but deviate from user intent, potentially leading to destructive security consequences. Furthermore, we observe that this vulnerability readily transfers to popular agent frameworks, while existing prompt-based mitigations remain largely ineffective. We call on the community to recognize this emerging vulnerability and develop effective mitigation methods to uphold the safety and reliability of the LLM system.

Large language model agents equipped with persistent memory are vulnerable to memory poisoning attacks, where adversaries inject malicious instructions through query only interactions that corrupt the agents long term memory and influence future responses. Recent work demonstrated that the MINJA (Memory Injection Attack) achieves over 95 % injection success rate and 70 % attack success rate under idealized conditions. However, the robustness of these attacks in realistic deployments and effective defensive mechanisms remain understudied. This work addresses these gaps through systematic empirical evaluation of memory poisoning attacks and defenses in Electronic Health Record (EHR) agents. We investigate attack robustness by varying three critical dimensions: initial memory state, number of indication prompts, and retrieval parameters. Our experiments on GPT-4o-mini, Gemini-2.0-Flash and Llama-3.1-8B-Instruct models using MIMIC-III clinical data reveal that realistic conditions with pre-existing legitimate memories dramatically reduce attack effectiveness. We then propose and evaluate two novel defense mechanisms: (1) Input/Output Moderation using composite trust scoring across multiple orthogonal signals, and (2) Memory Sanitization with trust-aware retrieval employing temporal decay and pattern-based filtering. Our defense evaluation reveals that effective memory sanitization requires careful trust threshold calibration to prevent both overly conservative rejection (blocking all entries) and insufficient filtering (missing subtle attacks), establishing important baselines for future adaptive defense mechanisms. These findings provide crucial insights for securing memory-augmented LLM agents in production environments.

Large Reasoning Models (LRMs) improve performance, reliability, and interpretability by generating explicit chain-of-thought (CoT) reasoning, but this transparency introduces a serious privacy risk: intermediate reasoning often leaks personally identifiable information (PII) even when final answers are sanitized. We study how to induce privacy-first reasoning, where models reason without exposing sensitive information, using deployable interventions rather than post-hoc redaction. We introduce PII-CoT-Bench, a supervised dataset with privacy-aware CoT annotations, and a category-balanced evaluation benchmark covering realistic and adversarial leakage scenarios. Our results reveal a capability-dependent trend: state-of-the-art models benefit most from prompt-based controls, whereas weaker models require fine-tuning to achieve meaningful leakage reduction. Across models and categories, both approaches substantially reduce PII exposure with minimal degradation in utility, demonstrating that private reasoning can be achieved without sacrificing performance. Overall, we show that private CoT reasoning can be achieved with minimal utility loss, providing practical guidance for building privacy-preserving reasoning systems.

Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, especially in multi-turn scenarios where harmful intent can be subtly injected across the conversation to produce nefarious outcomes. While single-turn attacks have been extensively explored, adaptability, efficiency and effectiveness continue to remain key challenges for their multi-turn counterparts. To address these gaps, we present PLAGUE, a novel plug-and-play framework for designing multi-turn attacks inspired by lifelong-learning agents. PLAGUE dissects the lifetime of a multi-turn attack into three carefully designed phases (Primer, Planner and Finisher) that enable a systematic and information-rich exploration of the multi-turn attack family. Evaluations show that red-teaming agents designed using PLAGUE achieve state-of-the-art jailbreaking results, improving attack success rates (ASR) by more than 30% across leading models in a lesser or comparable query budget. Particularly, PLAGUE enables an ASR (based on StrongReject) of 81.4% on OpenAI's o3 and 67.3% on Claude's Opus 4.1, two models that are considered highly resistant to jailbreaks in safety literature. Our work offers tools and insights to understand the importance of plan initialization, context optimization and lifelong learning in crafting multi-turn attacks for a comprehensive model vulnerability evaluation.

Generative AI leaderboards are central to evaluating model capabilities, but remain vulnerable to manipulation. Among key adversarial objectives is rank manipulation, where an attacker must first deanonymize the models behind displayed outputs -- a threat previously demonstrated and explored for large language models (LLMs). We show that this problem can be even more severe for text-to-image leaderboards, where deanonymization is markedly easier. Using over 150,000 generated images from 280 prompts and 19 diverse models spanning multiple organizations, architectures, and sizes, we demonstrate that simple real-time classification in CLIP embedding space identifies the generating model with high accuracy, even without prompt control or historical data. We further introduce a prompt-level separability metric and identify prompts that enable near-perfect deanonymization. Our results indicate that rank manipulation in text-to-image leaderboards is easier than previously recognized, underscoring the need for stronger defenses.

Memristive crossbar arrays enable in-memory computing by performing parallel analog computations directly within memory, making them well-suited for machine learning, neural networks, and neuromorphic systems. However, despite their advantages, non-volatile memristors are vulnerable to security threats (such as adversarial extraction of stored weights when the hardware is compromised. Protecting these weights is essential since they represent valuable intellectual property resulting from lengthy and costly training processes using large, often proprietary, datasets. As a solution we propose two security mechanisms: Keyed Permutor and Watermark Protection Columns; where both safeguard critical weights and establish verifiable ownership (even in cases of data leakage). Our approach integrates efficiently with existing memristive crossbar architectures without significant design modifications. Simulations across 45nm, 22nm, and 7nm CMOS nodes, using a realistic interconnect model and a large RF dataset, show that both mechanisms offer robust protection with under 10% overhead in area, delay and power. We also present initial experiments employing the widely known MNIST dataset; further highlighting the feasibility of securing memristive in-memory computing systems with minimal performance trade-offs.

Radio frequency fingerprint identification (RFFI) distinguishes wireless devices by the small variations in their analog circuits, avoiding heavy cryptographic authentication. While deep learning on spectrograms improves accuracy, models remain vulnerable to copying, tampering, and evasion. We present a stronger RFFI system combining watermarking for ownership proof and anomaly detection for spotting suspicious inputs. Using a ResNet-34 on log-Mel spectrograms, we embed three watermarks: a simple trigger, an adversarially trained trigger robust to noise and filtering, and a hidden gradient/weight signature. A convolutional Variational Autoencoders (VAE) with Kullback-Leibler (KL) warm-up and free-bits flags off-distribution queries. On the LoRa dataset, our system achieves 94.6% accuracy, 98% watermark success, and 0.94 AUROC, offering verifiable, tamper-resistant authentication.

AI web agents use Internet resources at far greater speed, scale, and complexity -- changing how users and services interact. Deployed maliciously or erroneously, these agents could overload content providers. At the same time, web agents can bypass CAPTCHAs and other defenses by mimicking user behavior or flood authentication systems with fake accounts. Yet providers must protect their services and content from denial-of-service attacks and scraping by web agents. In this paper, we design a framework that imposes tunable costs on agents before providing access to resources; we call this Web Agent Throttling. We start by formalizing Throttling Gates as challenges issued to an agent that are asymmetric, scalable, robust, and compatible with any agent. Focusing on a common component -- the language model -- we require the agent to solve reasoning puzzles, thereby incurring excessive token-generation costs. However, we find that using existing puzzles, e.g., coding or math, as throttling gates fails to satisfy our properties. To address this, we introduce rebus-based Reasoning Gates, synthetic text puzzles that require multi-hop reasoning over world knowledge (thereby throttling an agent's model). We design a scalable generation and verification protocol for such reasoning gates. Our framework achieves computational asymmetry, i.e., the response-generation cost is 9.2x higher than the generation cost for SOTA models. We further deploy reasoning gates on a custom website and Model Context Protocol (MCP) servers and evaluate with real-world web agents. Finally, we discuss the limitations and environmental impact of real-world deployment of our framework.

We show that Web and Research Agents (WRAs) -- language-model-based systems that investigate complex topics on the Internet -- are vulnerable to inference attacks by passive network observers. Deployment of WRAs \emph{locally} by organizations and individuals for privacy, legal, or financial purposes exposes them to DNS resolvers, malicious ISPs, VPNs, web proxies, and corporate or government firewalls. However, unlike sporadic and scarce web browsing by humans, WRAs visit $70{-}140$ domains per each request with a distinct timing pattern creating unique privacy risks. Specifically, we demonstrate a novel prompt and user trait leakage attack against WRAs that only leverages their network-level metadata (i.e., visited IP addresses and their timings). We start by building a new dataset of WRA traces based on real user search queries and queries generated by synthetic personas. We define a behavioral metric (called OBELS) to comprehensively assess similarity between original and inferred prompts, showing that our attack recovers over 73\% of the functional and domain knowledge of user prompts. Extending to a multi-session setting, we recover up to 19 of 32 latent traits with high accuracy. Our attack remains effective under partial observability and noisy conditions. Finally, we discuss mitigation strategies that constrain domain diversity or obfuscate traces, showing negligible utility impact while reducing attack effectiveness by an average of 29\%.