TrajGuard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense

Existing jailbreak defense paradigms primarily rely on static detection of prompts, outputs, or internal states, often neglecting the dynamic evolution of risk during decoding. This oversight leaves risk signals embedded in decoding trajectories underutilized, constituting a critical blind spot in current defense systems. In this work, we empirically demonstrate that hidden states in critical layers during the decoding phase carry stronger and more stable risk signals than input jailbreak prompts. Specifically, the hidden representations of tokens generated during jailbreak attempts progressively approach high-risk regions in the latent space. Based on this observation, we propose TrajGuard, a training-free, decoding-time defense framework. TrajGuard aggregates hidden-state trajectories via a sliding window to quantify risk in real time, triggering a lightweight semantic adjudication only when risk within a local window persistently exceeds a threshold. This mechanism enables the immediate interruption or constraint of subsequent decoding. Extensive experiments across 12 jailbreak attacks and various open-source LLMs show that TrajGuard achieves an average defense rate of 95%. Furthermore, it reduces detection latency to 5.2 ms/token while maintaining a false positive rate below 1.5%. These results confirm that hidden-state trajectories during decoding can effectively support real-time jailbreak detection, highlighting a promising direction for defenses without model modification.

LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation

As large language models (LLMs) scale, their inference incurs substantial computational resources, exposing them to energy-latency attacks, where crafted prompts induce high energy and latency cost. Existing attack methods aim to prolong output by delaying the generation of termination symbols. However, as the output grows longer, controlling the termination symbols through input becomes difficult, making these methods less effective. Therefore, we propose LoopLLM, an energy-latency attack framework based on the observation that repetitive generation can trigger low-entropy decoding loops, reliably compelling LLMs to generate until their output limits. LoopLLM introduces (1) a repetition-inducing prompt optimization that exploits autoregressive vulnerabilities to induce repetitive generation, and (2) a token-aligned ensemble optimization that aggregates gradients to improve cross-model transferability. Extensive experiments on 12 open-source and 2 commercial LLMs show that LoopLLM significantly outperforms existing methods, achieving over 90% of the maximum output length, compared to 20% for baselines, and improving transferability by around 40% to DeepSeek-V3 and Gemini 2.5 Flash.

Launching a Robust Backdoor Attack under Capability Constrained Scenarios

As deep neural networks continue to be used in critical domains, concerns over their security have emerged. Deep learning models are vulnerable to backdoor attacks due to the lack of transparency. A poisoned backdoor model may perform normally in routine environments, but exhibit malicious behavior when the input contains a trigger. Current research on backdoor attacks focuses on improving the stealthiness of triggers, and most approaches require strong attacker capabilities, such as knowledge of the model structure or control over the training process. These attacks are impractical since in most cases the attacker's capabilities are limited. Additionally, the issue of model robustness has not received adequate attention. For instance, model distillation is commonly used to streamline model size as the number of parameters grows exponentially, and most of previous backdoor attacks failed after model distillation; the image augmentation operations can destroy the trigger and thus disable the backdoor. This study explores the implementation of black-box backdoor attacks within capability constraints. An attacker can carry out such attacks by acting as either an image annotator or an image provider, without involvement in the training process or knowledge of the target model's structure. Through the design of a backdoor trigger, our attack remains effective after model distillation and image augmentation, making it more threatening and practical. Our experimental results demonstrate that our method achieves a high attack success rate in black-box scenarios and evades state-of-the-art backdoor defenses.