Another 233papers do not have category tags yet and are not included in filtered results.
提出回答端后门，让模型自生成触发词以在干净有害请求下绕过拒答
Abstract论文提出了多轮对话中回答端自生成触发词绕过安全拒绝的后门攻击，指出输入端防护存在盲区。
评测多轮对话中 GradSafe 越狱检测的可分性与脆弱性
作者评估梯度越狱检测从单条提示词迁到多轮对话后还剩多少可分性，方法是用 Context Window Scanner 把 GradSafe 原封不动地套到用户轮窗口上。
提出条件触发的爆炸提示词，并设计摄入期检测器 DeFuse
论文针对大语言模型智能体面临的间接提示注入威胁，研究了利用条件句式实现时间分离的爆炸提示词。
提出 MAFIA，以探测放置与事实伪装对带审计 Agent 做查询式记忆投毒
提出 CFD 跨会话分解攻击，利用工件来源缺口绕过工具型 Agent 的本地检查
CFD针对工具型 LLM 智能体：利用 artifact 来源不被跟踪这一部署失败模式，把有害目标拆开后延迟组合。
构建 MultiTurnPSB，评测四轮医疗越狱与分类器输入防御
MultiTurnPSB 针对患者端医疗大模型在持续交互中的安全风险，构建了包含三种攻击模态的四轮对抗评测基准。
Propose a structured JSON backdoor that triggers malicious actions in robot control via a poisoned LoRA
Abstract推理层后门传不到执行。
Using ADVERSA to measure guardrail degradation and reviewer reliability under multi-turn attacks
ADVERSA is an automated multi-turn red-teaming measurement framework that records guardrail behavior as per-turn compliance trajectories and simultaneously measures whether reviewers agree. The problem it addresses: single-turn binary pass/fail cannot show how safety properties change under sustained adversarial pressure, nor can it separate gradual refusal consolidation from stable scores. The authors argue that real-world adversaries rewrite and persist over multiple turns.
提出多轮多模态越狱，并用 FragGuard 拦截有害响应