Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 35 papers Another 232 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
提出 ReSA，从对齐模型行为恢复安全奖励并在解码时反转
Published Oct 6
Evaluated models Llama3.1-8B-Instruct , Gemma-7B-Instruct , Qwen2.5-7B-Instruct and others9 Abstract ReSA 用最大熵 IRL 从对齐模型行为恢复代理安全奖励，并在解码时反转以诱导有害输出。ReSA 是一种对抗攻击：从对齐模型的可观察行为恢复代理安全奖励，再在解码时把该奖励反过来用。 。
提出 CRIME 框架，组合良性 Agent 技能诱发恶意行为
Published Oct 5
Evaluated models DeepSeek-V4-Flash , GPT-4o-mini , GPT-4o and others6 Abstract 良性技能经 DCA 或 ACA 组合后可产生恶意环境后果，单独审查对不上。CRIME 把公共技能库中各自通过审查的技能配成两两组合，看联合执行能否实现指定恶意任务。 。
评测多轮对话中 GradSafe 越狱检测的可分性与脆弱性
Published Sep 29
Evaluated models Llama 3.1 8B Instruct , Qwen2.5-7B-Instruct Abstract CWS 显示不安全梯度在真实多轮流量中变弱，且随攻击族和模型而变。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。
提出能力洗钱攻击，拆分任务向对齐模型套取能力
Published Sep 14
Evaluated models GPT-5.5 , Claude Opus 4.8 , Grok-4.3 and others7 Abstract 作者提出了能力洗钱攻击，发现未对齐编排器分解咨询对齐模型可恢复能力差距，指出需建立跨会话防御。
提出 MIPC 与 CFA，观察模型自拟测验后接受开发者身份
Published Sep 3
Evaluated models ChatGPT , Claude , Qwen3.7-Plus and others5 Abstract 自拟测验可造成对话层虚假认证，但所测授权边界并未因此改变。这项工作把“模型自己出题、自己批改、再宣布身份已核验”定义成对话层认证失败，并与后端提权分开。
提出多轮越狱框架 BLUEPRINT，用世界观模拟与心理因素诱导有害输出
Published Sep 2
Evaluated models Qwen3-Next-80B , DeepSeek-V3.2 , GLM-4.7 and others6 Abstract BLUEPRINT 解耦多轮越狱策略与情境模拟，揭示任务具象化是摆脱拒绝的核心机制，为多轮安全防御提供新视角。BLUEPRINT 是一个将因子化社会影响力策略空间与跨轮世界观模拟解耦的多轮大语言模型安全评测框架。
提出面向意图的多轮 LLM 越狱分类体系
Published Aug 2
Evaluated models Llama-3 , Llama-3.3-70B , DeepSeek-V3 and others11 Abstract 作者用四维意图分类整理多轮越狱，并称组织方式影响效果与观察范围。这是一篇关于多轮 LLM 越狱的 SoK：按有害意图在交互中的组织位置分类，而不是按单步提示技巧分类。
提出生物红队模型 Intern-BioBreaker，多轮越狱诱导模型输出可操作生物危害
Published Jul 20
Evaluated models DeepSeek-V3 , Kimi-K2.6 , DeepSeek-V4-Pro and others14 Abstract 论文通过构建专用红队模型评估前沿大模型的生物安全漏洞与物理转化风险。
提出工作流级越狱，让编程 Agent 在开发任务中写入有害示例
Published Jul 5
Evaluated models Claude Sonnet 4.6 , Claude Haiku 4.5 , Gemini 3.1 Pro and others4 Abstract 作者报告 Copilot 后端会在编码工作流中生成其在单轮条件下通常拒绝的有害答案。这是一项生产 IDE 编码智能体的工作流越狱研究 ，受测对象是 Copilot 的四个模型后端，而非评测流水线中的名义目标。
提出 SMT 框架，用模拟审核轨迹越狱函数调用 LLM
Published Jul 1
Evaluated models GPT-4o , GPT-5.4 , Qwen3-Max and others7 提出 OTTER 词元替换越狱，绕过表面毒性审核诱导有害回答
Published Jun 19
Evaluated models gpt-4-turbo , gpt-4o , gpt-4o-mini and others4 Abstract 论文展示了表面毒性与恶意意图的解耦现象，通过标准 API 实现黑盒越狱测试，并提出闭环分类器加固建议。
提出 AdvGRPO，用 GRPO 闭环共训多轮越狱攻击者与防御者
Published Jun 9
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-14B-Instruct , Qwen3.5-9B and others7 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
Published Jun 8
Evaluated models GPT-4.1 , GPT-5-nano , Claude-3.5-Sonnet and others6 Abstract CFD 用无指令 artifact 把越狱拆到跨会话。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
Published May 30
Evaluated models GPT-4.1-mini , Claude Sonnet 4.5 Abstract 多轮攻击加剧医疗模型风险，分类器干预有效但受高误报率制约。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。
提出 TRACE 框架，用分解与可演化多轮策略诱导智能体执行有害工作流
Published May 29
Evaluated models Codex , Claude Code , GPT-5.5 and others6 提出 Persona Attack，分步写入对话记忆指令以越狱
Published May 29
Evaluated models GPT-4o , Llama-3.2-3B-Instruct , ChatGPT-4o and others5 提出多轮意图欺骗越狱 iDecep，用良性意图伪装引出有害内容
Published Apr 27
Evaluated models GPT-4o , Gemini-2.5-Flash , Claude-Sonnet-4.5 and others4 Abstract iDecep 用意图欺骗做多轮越狱，在 GPT-5 与 Claude-Sonnet-4.5 上同时测到直接成功和 para 成功。iDecep 是一种黑盒多轮越狱：用表面良性的意图维持与恶意目标贴近的对话，并单独计入 safe completion 替代内容中的有害信息。 。
Proposes the STAR multi-turn role-playing framework, which bypasses safety alignment through dialogue state evolution
Published Mar 15
Evaluated models GPT-4o , Claude 3.5 Sonnet , Gemini 2.0-Flash and others5 Using ADVERSA to measure guardrail degradation and reviewer reliability under multi-turn attacks
Published Mar 10
Evaluated models Claude Opus 4.6 , Gemini 3.1 Pro , GPT-5.2 Abstract ADVERSA measures multi-turn guardrail changes using per-turn compliance trajectories and three-reviewer consensus.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.
Propose Blindfold, jailbreaking embodied LLMs with action-level manipulation
Published Mar 2
Evaluated models Llama-3.1-8B , DeepSeek-R1-14B , Gemma-3-27B and others8 Abstract Blindfold jailbreaks embodied LLMs with action-level proxy planning, and the authors claim it can bypass semantic checks and is effective in both simulation and real-world settings.Blindfold is a no-box attack against embodied LLM planning modules, rewriting malicious intent into seemingly benign action sequences.