Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 15 papers Another 265 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.
Trace attention to identify harmfulness to the MLP refusal write pathway and propose Sentinel
Published Oct 9
Evaluated models Llama-3.1-8B-Instruct , Qwen2.5-7B-Instruct , Qwen3-8B-Instruct and others18 Abstract Attention updates are more readable for harmfulness, MLP updates write refusal, and Sentinel routes alerts using two-path signals.The authors reframe jailbreaking and excessive refusal of aligned models as a pathway problem on the residual stream from recognition to writing, and accordingly provide conditional reminders before decoding.
提出免训练框架 SED，将有害智能体轨迹蒸馏为可检索安全策略
Published Sep 29
Evaluated models DeepSeek-V4-Flash , GLM-5.2 , Kimi K3 提出 Safety Operator，用谱优化调节安全指令表达强度
Published Sep 29
Evaluated models Gemma 3 1B , Gemma 3 4B 提出工具型 Agent 的自适应攻击 DART 与预执行防御 SAGE
Published Sep 28
Evaluated models GPT-5.6 Luna , GPT-5.6 Terra , Gemini 3.8 Flash and others5 Abstract 论文提出状态控制框架 SEAD、攻击 DART 与防御 SAGE，实验显示执行反馈与环境探测在攻防交互中具有关键作用。
提出 Skilder，按角色渐进下发企业工具并确定性拦截越权调用
Published Sep 24
Evaluated models Claude Haiku 4.5 , Qwen 3.5 122B , Gemma 4 31B and others6 Abstract Skilder 用单 MCP 服务器按角色交付工具并硬拦截。Skilder 把企业工具型智能体的能力按角色打包，经一个 MCP 服务器做渐进发现，使工具范围由服务器而不是提示词来执行。
构建 APort Vault 基准，对比支付智能体模型拒答与确定性授权
Published Sep 19
Evaluated models Claude Fable 5.1 , Claude Sonnet 5 , Claude Haiku 4.5 and others14 提出 FLOWSEAL，用工具级信息流控制阻断智能体隐私泄露
Published Sep 12
Evaluated models DeepSeek-V3.2 , GPT-4.1-mini , Gemini-Flash-2.5 Abstract 作者分析大模型智能体隐私检查与攻击面重合问题，提出外部拦截防御 FLOWSEAL 并验证有效性。
用后验重加权解释多模态上下文越狱，并提出风险门控的良性示例注入防御
Published Sep 8
Evaluated models Qwen3-VL-2B , Qwen3-VL-4B , Qwen3-VL-8B and others8 Abstract 后验重加权解释上下文越狱的缩放，并支持风险门控的良性示例防御。Zhang 等把多模态 ICL 越狱写成对齐模型在安全与有害隐模式之间的推理时后验漂移，并据此做风险门控防御。
为冻结模型与目标域自动搜索可部署的 Agent 安全 harness
Published Sep 5
Evaluated models Sonnet 4.6 , GLM-5 , Kimi-K2.5 and others11 提出 AlcaTRAz，用锚定字符规则树改写提示以防御越狱
Published Sep 3
Evaluated models LLaMA 2 , LLaMA 3.1 , LLaMA 3.2 and others14 提出 SafeEvolve，用轨迹证据协同演化安全 harness 与策略
Published Sep 3
Evaluated models Qwen3.5-4B , Qwen3-4B-Instruct-2507 , Qwen3.5-1.7B and others4 Abstract SafeEvolve 用轨迹证据协同更新安全 harness 和策略，在两个 4B 模型的智能体安全基准上改变安全与效用。SafeEvolve 把已完成的 on-policy 轨迹变成安全 harness 的有界更新，并用两阶段 SFT-RL 把这些更新内化进工具使用策略。
提出 Skill-as-API，隔离多智能体协作中的 skill 正文与系统提示词
Published Sep 1 Abstract 协议把 skill 正文留在所有者侧，四层限制发现内容、权限期限和注入拼接。Skill-as-API 是面向智能体软件工程的协调协议，用来在跨组织调用时把 skill 的代码和系统提示词留在所有者进程。作者针对三条威胁：经通道或命名空间抽取专有方法，用注入让被调用模型输出系统提示词，以及协作结束后权限继续保留。
提出 GIFT，用加密与 GPU 信息流跟踪隔离共享 LLM 服务中的用户数据
Published Aug 26
Evaluated models Qwen2.5-14B , Qwen2.5-32B , Qwen2.5-72B and others6 Abstract GIFT 以加密隔离加 GPU 信息流规则，在 vLLM 与 DistServe 上强制用户数据隔离，并报告了吞吐开销。GIFT 是作者为共享 LLM serving 做的用户数据隔离系统，并有启用机密计算的 GIFT-CC。作者在结论中称它是该场景的首个 IFT 设计。
分析少样本示范对 RoP 与 ToP 提示防御抗越狱效果的影响
Published Feb 4
Evaluated models Pangu-Embedded-1B , Pangu-Embedded-7B , Pangu-Embedded-7B-Think and others9 Abstract few-shot 可抬高 RoP、压低 ToP，但表内有反号，Think 模型更低。作者评估 few-shot 示范如何改变 RoP 与 ToP 对越狱的防御效果。
提出 DASE，以对抗自演化知识库在推理时加固角色扮演抗越狱
Published Jan 29
Evaluated models Kimi-K2-Instruct , GPT-5.2 Abstract DASE 用双循环知识库，在两家闭源模型上同时提高保真分与拒答率。DASE 是给专有闭源角色扮演智能体用的训练自由防御，目标是人设保真与安全拒答一起成立，而不是改模型权重。问题在于人设沉浸，尤其负面人设，会压过原有护栏。SFT、数据筛选和正则又跟不上新人设与新攻击，也不能用于没有梯度的 black-box API。
You have reached the end