Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 28 papers Another 261 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.
Proposes the Cancerous backdoor, which progressively corrupts reasoning using invisible stage tokens.
Published Oct 9
Evaluated models Llama-3.1-8B-Instruct , Qwen2.5-7B-Instruct , Mistral-7B-Instruct-v0.3 Abstract Hidden tokens plus a stage controller make the backdoor progressively corrupt reasoning upon triggering, and the authors believe that looking only at the answer will miss it.Cancerous Backdoor Attack changes LLM backdoors from one-time activation to multi-stage reasoning corruption that accumulates progressively with triggers.
Uses premise permutation and hidden-layer tracing to audit whether legal chain-of-thought truly relies on the cited legal provisions.
Published Oct 9
Evaluated models Qwen3-8B , Qwen3-14B , Llama-3.1-8B and others7 Abstract Using counterfactual substitution of authorities and internal commitment tracing, the authors find that legal CoT rulings depend on authorities only to a limited extent and are vulnerable to prompt injection manipulation.This study conducts an empirical audit of the causal truthfulness of legal citations in chain-of-thought reasoning of large language models in the legal domain.
提出 ACTR，只更新安全思考神经元以对齐推理与回答的安全类别
Published Sep 29
Evaluated models Qwen3-8B , Gemma4-12B-it 构建 TriggerBench 并训练 ThinkingGuard，检测图文隐式风险
Published Sep 29
Evaluated models ThinkingGuard（Qwen3-VL-8B-Instruct） Abstract TriggerBench 隔离触发条件，ThinkingGuard 用分步归因在隐式风险检测上取得作者报告的较高平均分。这篇工作研究多模态隐式风险检测 ：图文各自良性，风险只在关键元素与触发元素交互时出现。
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 EOPSA，用 TRR 截断与安全类过滤做在策略自蒸馏安全对齐
Published Sep 28
Evaluated models Qwen3-1.7B , Qwen3-4B , Qwen3-14B and others5 提出推理通道预填与输出前缀组合攻击以越狱推理模型
Published Sep 24
Evaluated models Gemini 3 Flash Preview , DeepSeek V4 Flash , Claude Haiku 4.5 分析量化、剪枝与 KV 缓存压缩对推理模型越狱鲁棒性的影响
Published Sep 20
Evaluated models QwQ-32B , DeepSeek-R1-Distill-Qwen-32B , DeepSeek-R1-Distill-Llama-8B Abstract 效率压缩常使 LRM 越狱 HR 下降，作者将其归因于推理退化而非对齐增强。作者分析量化、剪枝和 KV 缓存压缩如何同时改变 LRM 的显存、推理和越狱脆弱性。 。
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 提出 plan injection 攻击，植入良性伪装计划以绕过思维链监视器
Published Sep 15
Evaluated models Qwen2.5-Coder-7B-Instruct , Qwen3-8B , DeepSeek-R1-0528 and others6 Abstract 论文指出模型会将外部上下文中的伪装计划改写为自身推理，使思维链监视器失效甚至反向合理化恶意行为。
系统梳理智能体执行链路中的越狱攻击与防御并比较推理时防御
Published Sep 11
Evaluated models Qwen3.5-4B , Llama-3.1-8B-Instruct Abstract SoK 将越狱重放到智能体流水线，并用两种模型检验回复层防御的边际与代价。Mia 等人把越狱安全从单轮回复扩展到智能体的规划、记忆、工具和智能体间通信，并问既有结论在更强原生对齐下是否还成立。
提出任意字母置换密码越狱，无需微调即可诱导有害输出
Published Sep 9
Evaluated models Claude Sonnet 4 , Claude Sonnet 4.5 , Gemini 3 Flash (preview) and others7 Abstract 前沿大模型无需微调即可通过提示词学会字母置换密码，进而绕过分类器与对齐，作者呼吁关注能力提升带来的安全风险。
揭示推理加深引发的对齐坍塌，用 Reasoning Trap 放大越狱
Published Sep 8
Evaluated models Qwen3-8B , Qwen3-14B , Llama2-7B and others8 Abstract 更深推理提升题目准确率，同时提高扰动下的相对损失。作者研究扩展推理是否以对齐稳健性为代价，提出 ALR、RT 与 RRA。 。
提出隐蔽后门攻击 AKRASIA，在推理代码模型提示中植入触发器并伪装思维链
Published Sep 1
Evaluated models Claude-Sonnet-5 , GPT-5.5 , Gemini-3.5-flash and others7 Abstract 论文针对推理代码大模型提出了隐蔽推断期后门攻击 AKRASIA，揭示了模型思维链可被利用进行欺骗性伪装的安全隐患。
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 RolePlay 框架，用人设条件放大 LLM 单次推理开销
Published Jul 29
Evaluated models DeepSeek-V4-Pro , Gemini-3.5-Flash , Qwen-3.5-Plus and others5 提出嵌入推断攻击 EIA，在黑盒 IR 中识别所用嵌入模型
Published Jul 1
Evaluated models e5-small , e5-small-v2 , e5-base and others19 Abstract EIA 用无序文档集的差异从候选中识别嵌入模型，并给出 reranker 与阈值下的计数。EIA 是在黑盒 IR 上识别所用嵌入模型的攻击。 。
提出层级遗传算法 HGA，扰动题目逻辑结构诱发推理模型过度思考
Published May 13
Evaluated models DeepSeek-R1 (671B) , Qwen3-Thinking , GPT-o3 and others5 Abstract HGA 在黑盒推理模型上用逻辑扰动拉长输出，作者称可从小代理迁移，查询成本是其局限。
提出复合越狱，叠对比结构、权威人设与自评饱和指令层级
Published Apr 3
Evaluated models OpenAI gpt-oss-20b 提出 VROP，把有害意图拆成良性视觉 gadget 越狱视觉语言模型
Published Mar 10
Evaluated models GPT-4o , Claude 3.7 Sonnet , GLM-4V-Plus and others7 Abstract VROP 以良性视觉 gadget 与控制流提示词做组合越狱。VROP 是 black-box 多模态越狱框架：不在输入中放显式恶意触发，而把有害目标拆成语义正交的良性视觉 gadget，再用控制流提示词把合成留到后期推理。