Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
Another 233 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.
评测文生图模型的有害内容生成能力与审核检测缺口
Published Oct 5
Evaluated models SDXL , Stable Diffusion 3.5 Large , FLUX.1-dev and others12 提出 Persona Guardrail，强制客服智能体守住功能边界
Published Oct 2
Evaluated models Qwen3-30B-A3B , Qwen3-VL-30B , Qwen3-8B and others5 Abstract Persona Guardrail 用功能边界做同步输入输出校验，并在 PAGE 与生产延迟预算下给出评测。Persona Guardrail 是客服向智能体的同步运行时防御：先写明功能边界，再决定放行或拦截，而不是只检测某一攻击类。
提出用 API 推理摘要重建轨迹再蒸馏加强化学习，攻破蒸馏防御
Published Sep 29
Evaluated models Qwen2.5-0.5B , Qwen2.5-1.5B , Qwen2.5-3B and others12 Abstract 更现实的蒸馏攻击是蒸馏后再 RL。这篇工作修订蒸馏攻击的威胁模型：攻击者经合法 API 收集闭源推理数据，先蒸馏自有模型，再做强化学习。作者认为只在蒸馏后评估防御，会漏掉后续 RL，从而高估防御。
提出 ARIA，在冻结权重上用 SAE 做推理时遗忘门控
Published Sep 15
Evaluated models DeepSeek-R1-Distill-Qwen-1.5B , Gemma-3-1B-it 利用加密思维块的跨模型重放，诱导弱模型逐字转写专有推理链
Published Aug 11
Evaluated models Claude Opus 4.8 , Claude Sonnet 4.6 , Claude Haiku 4.5 and others13 Abstract 论文揭示了客户端加密思维块跨模型互换漏洞，提出了低成本提取方法并给出防御方案。
提出 RA-Det，利用鲁棒性不对称检测 AI 生成图像
Published Mar 2 Abstract RA-Det 把扰动下特征漂移差变成检测信号，平均 ACC 高于表中通用检测器，但不是每行最高。RA-Det 是合成图像检测器，用扰动下的表示稳定性差异代替外观痕迹。
提出 Staged Decoding 与指令跟随训练，减少推理轨迹隐私泄露
Published Feb 28
Evaluated models Qwen 3 1.7B , Qwen 3 4B , Qwen 3 8B and others6 Abstract 用一般指令跟随训练加分段解码，可降低轨迹隐私泄露，但数学效用下降。本文把大推理模型的隐私问题做成轨迹可控性：隐私指令是指令，轨迹不跟随就会把敏感上下文写进可被提示注入导出的文本。
提出推理时框架 STaR，抑制大推理模型推理轨迹中的敏感内容
Published Jan 14
Evaluated models DeepSeekR1-Distill-Llama-8B Abstract STaR 以推理时轨迹调控做 LRM unlearning，作者称在 R-TOFU 上遗忘与隐私指标更好。STaR 在推理阶段约束大推理模型的整段 CoT，使指定 forget set 中的敏感内容不出现在中间步骤或最终答案里，且不更新参数。问题来自 LRM 会把敏感知识写进多步推理，而答案级遗忘在 ZeroThink、LessThink 等解码下仍可能泄漏。STaR 先判断查询是否相关并检索遗忘片段，再加安全前缀。
Defenses arXiv 2312.06632 · Dec 12, 2023 提出 SciGuard 外部护栏与 SciMT 基准，控制化学科学 AI 误用
Published Dec 12, 2023
Evaluated models SciGuard (GPT-4) , SciGuard (Gemini) , SciGuard (Mixtral) and others5 Abstract SciGuard 以外部智能体控制化学 AI 误用。这篇工作把化学科学 AI 的误用控制做成外部智能体，而不是改模型权重。作者担心合成规划、毒性预测和科学语言模型被用来获取有害物质信息、规避监管或传播危险知识，同时合法工业用途又不能靠一律拒绝来处理。SciGuard 由 LLM 担任中介，把原则和指南放进短期记忆，把分子库、危害库、法规和交互历史放进可检索的长期记忆，再按 Chain of Thought 调用 Chemprop、LocalRetro 等工具，或用 RAG 补法规片段，最后回答、追问或拒绝。
You have reached the end