Fewer Weights, More Problems: A Practical Attack on LLM Pruning

Model pruning, i.e., removing a subset of model weights, has become a prominent approach to reducing the memory footprint of large language models (LLMs) during inference. Notably, popular inference engines, such as vLLM, enable users to conveniently prune downloaded models before they are deployed. While the utility and efficiency of pruning methods have improved significantly, the security implications of pruning remain underexplored. In this work, for the first time, we show that modern LLM pruning methods can be maliciously exploited. In particular, an adversary can construct a model that appears benign yet, once pruned, exhibits malicious behaviors. Our method is based on the idea that the adversary can compute a proxy metric that estimates how likely each parameter is to be pruned. With this information, the adversary can first inject a malicious behavior into those parameters that are unlikely to be pruned. Then, they can repair the model by using parameters that are likely to be pruned, effectively canceling out the injected behavior in the unpruned model. We demonstrate the severity of our attack through extensive evaluation on five models; after any of the pruning in vLLM are applied (Magnitude, Wanda, and SparseGPT), it consistently exhibits strong malicious behaviors in a diverse set of attack scenarios (success rates of up to $95.7\%$ for jailbreak, $98.7\%$ for benign instruction refusal, and $99.5\%$ for targeted content injection). Our results reveal a critical deployment-time security gap and underscore the urgent need for stronger security awareness in model compression.

BrokenMath: A Benchmark for Sycophancy in Theorem Proving with LLMs

Large language models (LLMs) have recently shown strong performance on mathematical benchmarks. At the same time, they are prone to hallucination and sycophancy, often providing convincing but flawed proofs for incorrect mathematical statements provided by users. This significantly limits the applicability of LLMs in theorem proving, as verification of these flawed proofs must be done manually by expert mathematicians. However, existing benchmarks that measure sycophancy in mathematics are limited: they focus solely on final-answer problems, rely on very simple and often contaminated datasets, and construct benchmark samples using synthetic modifications that create ill-posed questions rather than well-posed questions that are demonstrably false. To address these issues, we introduce BrokenMath, the first benchmark for evaluating sycophantic behavior in LLMs within the context of natural language theorem proving. BrokenMath is built from advanced 2025 competition problems, which are perturbed with an LLM to produce false statements and subsequently refined through expert review. Using an LLM-as-a-judge framework, we evaluate state-of-the-art LLMs and agentic systems and find that sycophancy is widespread, with the best model, GPT-5, producing sycophantic answers 29% of the time. We further investigate several mitigation strategies, including test-time interventions and supervised fine-tuning on curated sycophantic examples. These approaches substantially reduce, but do not eliminate, sycophantic behavior.

Constrained Decoding of Diffusion LLMs with Context-Free Grammars

Large language models (LLMs) have shown promising performance across diverse domains. Many practical applications of LLMs, such as code completion and structured data extraction, require adherence to syntactic constraints specified by a formal language. Yet, due to their probabilistic nature, LLM output is not guaranteed to adhere to such formal languages. Prior work has proposed constrained decoding as a means to restrict LLM generation to particular formal languages. However, existing works are not applicable to the emerging paradigm of diffusion LLMs, when used in practical scenarios such as the generation of formally correct C++ or JSON output. In this paper we address this challenge and present the first constrained decoding method for diffusion models, one that can handle formal languages captured by context-free grammars. We begin by reducing constrained decoding to the more general additive infilling problem, which asks whether a partial output can be completed to a valid word in the target language. This problem also naturally subsumes the previously unaddressed multi-region infilling constrained decoding. We then reduce this problem to the task of deciding whether the intersection of the target language and a regular language is empty and present an efficient algorithm to solve it for context-free languages. Empirical results on various applications, such as C++ code infilling and structured data extraction in JSON, demonstrate that our method achieves near-perfect syntactic correctness while consistently preserving or improving functional correctness. Importantly, our efficiency optimizations ensure that the computational overhead remains practical.

MathArena: Evaluating LLMs on Uncontaminated Math Competitions

The rapid advancement of reasoning capabilities in large language models (LLMs) has led to notable improvements on mathematical benchmarks. However, many of the most commonly used evaluation datasets (e.g., AIME 2024) are widely available online, making it difficult to disentangle genuine reasoning from potential memorization. Furthermore, these benchmarks do not evaluate proof-writing capabilities, which are crucial for many mathematical tasks. To address this, we introduce MathArena, a new benchmark based on the following key insight: recurring math competitions provide a stream of high-quality, challenging problems that can be used for real-time evaluation of LLMs. By evaluating models as soon as new problems are released, we effectively eliminate the risk of contamination. Using this framework, we find strong signs of contamination in AIME 2024. Nonetheless, evaluations on harder competitions, such as SMT 2025 -- published well after model release dates -- demonstrate impressive reasoning capabilities in top-performing models. MathArena is also the first benchmark for proof-writing capabilities. On USAMO 2025, even top models score below 25%, far behind their performance on final-answer tasks. So far, we have evaluated 30 models across five competitions, totaling 149 problems. As an evolving benchmark, MathArena will continue to track the progress of LLMs on newly released competitions, ensuring rigorous and up-to-date evaluation of mathematical reasoning.

Mind the Gap: A Practical Attack on GGUF Quantization

With the increasing size of frontier LLMs, post-training quantization has become the standard for memory-efficient deployment. Recent work has shown that basic rounding-based quantization schemes pose security risks, as they can be exploited to inject malicious behaviors into quantized models that remain hidden in full precision. However, existing attacks cannot be applied to more complex quantization methods, such as the GGUF family used in the popular ollama and llama.cpp frameworks. In this work, we address this gap by introducing the first attack on GGUF. Our key insight is that the quantization error -- the difference between the full-precision weights and their (de-)quantized version -- provides sufficient flexibility to construct malicious quantized models that appear benign in full precision. Leveraging this, we develop an attack that trains the target malicious LLM while constraining its weights based on quantization errors. We demonstrate the effectiveness of our attack on three popular LLMs across nine GGUF quantization data types on three diverse attack scenarios: insecure code generation ($\Delta$=$88.7\%$), targeted content injection ($\Delta$=$85.0\%$), and benign instruction refusal ($\Delta$=$30.1\%$). Our attack highlights that (1) the most widely used post-training quantization method is susceptible to adversarial interferences, and (2) the complexity of quantization schemes alone is insufficient as a defense.

Robust LLM Fingerprinting via Domain-Specific Watermarks

As open-source language models (OSMs) grow more capable and are widely shared and finetuned, ensuring model provenance, i.e., identifying the origin of a given model instance, has become an increasingly important issue. At the same time, existing backdoor-based model fingerprinting techniques often fall short of achieving key requirements of real-world model ownership detection. In this work, we build on the observation that while current open-source model watermarks fail to achieve reliable content traceability, they can be effectively adapted to address the challenge of model provenance. To this end, we introduce the concept of domain-specific watermarking for model fingerprinting. Rather than watermarking all generated content, we train the model to embed watermarks only within specified subdomains (e.g., particular languages or topics). This targeted approach ensures detection reliability, while improving watermark durability and quality under a range of real-world deployment settings. Our evaluations show that domain-specific watermarking enables model fingerprinting with strong statistical guarantees, controllable false positive rates, high detection power, and preserved generation quality. Moreover, we find that our fingerprints are inherently stealthy and naturally robust to real-world variability across deployment scenarios.

MixAT: Combining Continuous and Discrete Adversarial Training for LLMs

Despite recent efforts in Large Language Models (LLMs) safety and alignment, current adversarial attacks on frontier LLMs are still able to force harmful generations consistently. Although adversarial training has been widely studied and shown to significantly improve the robustness of traditional machine learning models, its strengths and weaknesses in the context of LLMs are less understood. Specifically, while existing discrete adversarial attacks are effective at producing harmful content, training LLMs with concrete adversarial prompts is often computationally expensive, leading to reliance on continuous relaxations. As these relaxations do not correspond to discrete input tokens, such latent training methods often leave models vulnerable to a diverse set of discrete attacks. In this work, we aim to bridge this gap by introducing MixAT, a novel method that combines stronger discrete and faster continuous attacks during training. We rigorously evaluate MixAT across a wide spectrum of state-of-the-art attacks, proposing the At Least One Attack Success Rate (ALO-ASR) metric to capture the worst-case vulnerability of models. We show MixAT achieves substantially better robustness (ALO-ASR < 20%) compared to prior defenses (ALO-ASR > 50%), while maintaining a runtime comparable to methods based on continuous relaxations. We further analyze MixAT in realistic deployment settings, exploring how chat templates, quantization, low-rank adapters, and temperature affect both adversarial training and evaluation, revealing additional blind spots in current methodologies. Our results demonstrate that MixAT's discrete-continuous defense offers a principled and superior robustness-accuracy tradeoff with minimal computational overhead, highlighting its promise for building safer LLMs. We provide our code and models at https://github.com/insait-institute/MixAT.

Finetuning-Activated Backdoors in LLMs

Finetuning openly accessible Large Language Models (LLMs) has become standard practice for achieving task-specific performance improvements. Until now, finetuning has been regarded as a controlled and secure process in which training on benign datasets led to predictable behaviors. In this paper, we demonstrate for the first time that an adversary can create poisoned LLMs that initially appear benign but exhibit malicious behaviors once finetuned by downstream users. To this end, our proposed attack, FAB (Finetuning-Activated Backdoor), poisons an LLM via meta-learning techniques to simulate downstream finetuning, explicitly optimizing for the emergence of malicious behaviors in the finetuned models. At the same time, the poisoned LLM is regularized to retain general capabilities and to exhibit no malicious behaviors prior to finetuning. As a result, when users finetune the seemingly benign model on their own datasets, they unknowingly trigger its hidden backdoor behavior. We demonstrate the effectiveness of FAB across multiple LLMs and three target behaviors: unsolicited advertising, refusal, and jailbreakability. Additionally, we show that FAB-backdoors are robust to various finetuning choices made by the user (e.g., dataset, number of steps, scheduler). Our findings challenge prevailing assumptions about the security of finetuning, revealing yet another critical attack vector exploiting the complexities of LLMs.

Type-Constrained Code Generation with Language Models

Large language models (LLMs) have achieved notable success in code generation. However, they still frequently produce uncompilable output because their next-token inference procedure does not model formal aspects of code. Although constrained decoding is a promising approach to alleviate this issue, it has only been applied to handle either domain-specific languages or syntactic language features. This leaves typing errors, which are beyond the domain of syntax and generally hard to adequately constrain. To address this challenge, we introduce a type-constrained decoding approach that leverages type systems to guide code generation. We develop novel prefix automata for this purpose and introduce a sound approach to enforce well-typedness based on type inference and a search over inhabitable types. We formalize our approach on a simply-typed language and extend it to TypeScript to demonstrate practicality. Our evaluation on HumanEval shows that our approach reduces compilation errors by more than half and increases functional correctness in code synthesis, translation, and repair tasks across LLMs of various sizes and model families, including SOTA open-weight models with more than 30B parameters.

Proof or Bluff? Evaluating LLMs on 2025 USA Math Olympiad

Recent math benchmarks for large language models (LLMs) such as MathArena indicate that state-of-the-art reasoning models achieve impressive performance on mathematical competitions like AIME, with the leading model, o3-mini, achieving scores comparable to top human competitors. However, these benchmarks evaluate models solely based on final numerical answers, neglecting rigorous reasoning and proof generation which are essential for real-world mathematical tasks. To address this, we introduce the first comprehensive evaluation of full-solution reasoning for challenging mathematical problems. Using expert human annotators, we evaluated several state-of-the-art reasoning models on the six problems from the 2025 USAMO within hours of their release. Our results reveal that all tested models struggled significantly, achieving less than 5% on average. Through detailed analysis of reasoning traces, we identify the most common failure modes and find several unwanted artifacts arising from the optimization strategies employed during model training. Overall, our results suggest that current LLMs are inadequate for rigorous mathematical reasoning tasks, highlighting the need for substantial improvements in reasoning and proof generation capabilities.