提出三类防御策略，有效阻止大模型被越狱攻击
Defending Large Language Models Against Jailbreak Exploits with Responsible AI Considerations
- 通过输入净化、重写和动态防护检测恶意提示
- 在推理时调整安全层向量，强化拒绝有害请求
- 用角色分工的智能体框架实现领域合规控制
大型语言模型（LLMs）仍易受越狱攻击影响，可绕过安全过滤并诱发有害或不道德行为。本文系统梳理了提示级、模型级和训练期干预三类现有防御方法，并提出三项新策略：第一，提示级防御框架通过净化、重写和自适应保护检测并中和恶意输入；第二，基于逻辑向量的引导防御在推理阶段对安全敏感层进行向量调控，增强拒绝行为；第三，领域特定智能体防御采用MetaGPT框架，通过结构化角色协作与领域遵从机制实现可控生成。在基准数据集上的实验显示，攻击成功率显著降低，尤其在智能体防御下实现完全抑制。研究揭示越狱攻击对大模型构成重大安全威胁，识别出关键干预节点，但防御策略常需在安全、性能与可扩展性间权衡。代码已开源。
原文摘要 · Abstract (English)
Large Language Models (LLMs) remain susceptible to jailbreak exploits that bypass safety filters and induce harmful or unethical behavior. This work presents a systematic taxonomy of existing jailbreak defenses across prompt-level, model-level, and training-time interventions, followed by three proposed defense strategies. First, a Prompt-Level Defense Framework detects and neutralizes adversarial inputs through sanitization, paraphrasing, and adaptive system guarding. Second, a Logit-Based Steering Defense reinforces refusal behavior through inference-time vector steering in safety-sensitive layers. Third, a Domain-Specific Agent Defense employs the MetaGPT framework to enforce structured, role-based collaboration and domain adherence. Experiments on benchmark datasets show substantial reductions in attack success rate, achieving full mitigation under the agent-based defense. Overall, this study highlights how jailbreaks pose a significant security threat to LLMs and identifies key intervention points for prevention, while noting that defense strategies often involve trade-offs between safety, performance, and scalability. Code is available at: https://github.com/Kuro0911/CS5446-Project
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。