Despite extensive safety alignment, Large Language Models (LLMs) remain vulnerable to jailbreak attacks.However, existing methods generally lack the capability for continuous learning and self-evolution from interactions, limiting the diversity and adaptability of attack strategies.To address this, we propose ASTRA, an automated framework capable of autonomously discovering, retrieving, and evolving attack strategies.ASTRA operates on a closed-loop "attack-evaluate-distillreuse" mechanism, which not only generates attack prompts but also automatically distills reusable strategies from every interaction.To systematically manage these strategies, we introduce a dynamic three-tier strategy library (Effective, Promising, and Ineffective) that categorizes strategies based on performance.This hierarchical memory mechanism enables the framework to enhance efficiency by leveraging successful patterns while optimizing the exploration space by avoiding known failures.Extensive experiments in a black-box setting demonstrate that ASTRA significantly outperforms existing baselines.

Cloud computing has revolutionized data center management by providing scalable and efficient resources for processing and data management. However, deploying containerd-based services in data centers presents significant challenges: (1) Active servers that are underutilized result in high energy consumption, necessitating optimization for energy efficiency; (2) Affinity requirements between services and servers must be considered to ensure appropriate deployments; (3) Quality of Service (QoS) requirements must be met, particularly to avoid performance interference when multiple services are deployed on the same server. To address these challenges, we propose a novel algorithm, Affinity-Interference Energy Deployment (AIED), based on Deep Reinforcement Learning (DRL). This algorithm strategically consolidates services onto fewer servers to optimize energy efficiency while adhering to stringent QoS and affinity constraints. By employing a demand-supply model to quantify QoS requirements and formulating the deployment challenge as a Markov Decision Process (MDP), our algorithm dynamically adapts to fluctuating demands and resource availability. Extensive simulations demonstrate that AIED significantly outperforms existing baseline strategies, reducing energy consumption while ensuring robust compliance with both QoS and affinity constraints.

ABSTRACT Edge computing systems are in great need of task scheduling due to resource constraints. However, existing scheduling algorithms typically optimize single objectives and lack adaptability to varying system conditions, failing to balance response time minimization during low workloads with queue balance maintenance under high workloads. This paper proposes an adaptive task scheduling algorithm based on Soft Actor‐Critic (SAC) with a novel workload‐aware reward mechanism, which automatically transitions between response time optimization and queue balance prioritization according to system load conditions. The whole scheduling problem is modeled as a Markov Decision Process (MDP), and a sliding window‐based performance evaluation framework is introduced to provide robust system assessment. Extensive experiments across multiple scenarios demonstrate that our method consistently achieves optimal response time across varying workload conditions, significantly outperforming traditional scheduling algorithms, while maintaining effective queue balance comparable to load‐based approaches under high workload scenarios.