Translation in progress Some entries still include text in the original language.
Oct 13 · Tuesday
All papers Found 27 papers Another 203 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 digital lock compilation, encrypting executable networks into host weights to hide arbitrary capabilities
Published Oct 10
Evaluated models Gemma 3 270M Abstract The authors use a digital lock to hide an executable capability and state that elicitation on the chemistry and parity tasks did not recover the capability.The authors construct a class of backdoors that encrypt an executable network into the weights, used to discuss the limits of white-box elicitation audits under supply chain attacks.
Retest and extend prompt injection against AI agents, and use Agent Card deception to manipulate A2A selection
Published Oct 8
Evaluated models Ollama Mistral 0.9.5 , llama3-8b-8192 , gemini-2.0-flash-lite and others5 Abstract Single-agent filtering and classification can be bypassed by format or wording, and A2A card selection is also biased toward names and descriptions.The authors retested prompt injection on their own LangChain, AutoGen, and A2A systems, and modified the Agent Card to see if choices shifted. The abstract states the experiments mainly highlight single-agent weaknesses while extending attacks to multi-agent.
提出 WebMirage，用局部对抗图像劫持视觉网页 Agent 的浏览器操作
Published Oct 7
Evaluated models LLaVA-v1.5-13B , LLaVA-v1.6-34B , MiniCPM-o-8B and others6 Abstract 论文提出了针对网络智能体的端到端视觉红队框架 WebMirage，揭示了从视觉定位到浏览器执行的安全脆弱性。WebMirage 是一个面向视觉定位网络智能体的流水线感知端到端红队攻击框架。
提出 PERSISTBD，发布前强化供应链后门使其经良性 SFT 与 RL 后仍外发凭据
Published Oct 6
Evaluated models Qwen2.5-Coder-3B-Instruct , Qwen2.5-Coder-7B-Instruct , Qwen3-Coder-30B-A3B-Instruct Abstract 总结了论文关于后门在 SFT/RL 中存活的发现、两因子机制、PERSISTBD 方法与核心数字。本文研究第三方 LLM 智能体在开发者进行良性监督微调（SFT）和强化学习（RL）后训练时的后门持久性风险。
提出只改 SAE 解码器的后门，在冻结语言模型上植入代码插入
Published Oct 5
Evaluated models Gemma 3 4B Instruct , Qwen3 1.7B , Llama 3.1 8B Instruct Abstract 只改解码器的 SAE 可带上代码插入后门，且若干常规质量指标变化不大。只改 SAE 解码器、冻结编码器与语言模型，即可把代码插入行为放进推理时的辅助组件。作者把这视为 SAE 供应链问题：checkpoint 一旦替换某一层激活，就参与后续计算，而不只是解释工具。
提出闭环激活引导，在扩散语言模型去噪过程中注入定向偏见
Published Oct 5
Evaluated models LLaDA-8B-Instruct , Dream-v0-Instruct-7B , LLaDA-MoE-7B-A1B-Instruct Abstract 通过去噪中间概率反馈动态调节残差流转向，证明扩散语言模型在推理服务栈中存在闭环偏见注入脆弱性。
提出核化激活引导 KAS，在推理时自适应转向激活以诱导越狱
Published Oct 1
Evaluated models Qwen2.5-7B-Instruct , Gemma-2-9B-IT , Llama-3.1-8B-Instruct and others4 Abstract KAS 用核梯度做局部自适应转向，并在线性核下恢复 DiM。Kernelized Activation Steering（KAS） 把推理时的激活转向写成核空间中的优化，使更新随当前激活相对源、目标参考集的位置而变化。
提出只训练或篡改多智能体潜在通信链路来提高有害遵从
Published Sep 30
Evaluated models Llama-3.2-3B-Instruct , Qwen2.5-3B-Instruct , Gemma-3-1B-IT and others7 Abstract 潜在链路本身即可改变冻结智能体系统的安全，奖励既能攻击也能修复。学习得到的潜在通信链路会改变由冻结、对齐智能体组成的系统是否遵从有害请求。 。作者研究三种拓扑：双智能体、三级顺序协作，以及双专家加汇总。
提出检索器后门攻击操纵智能体搜索，并用 DOU 逃避检测
Published Sep 26
Evaluated models E5 , Contriever-MSMARCO Abstract 论文展示了通过恶意检索器操纵智能体多轮搜索并利用诱饵重塑表征隐蔽后门的可行性与系统风险。
提出条件触发的爆炸提示词，并设计摄入期检测器 DeFuse
Published Sep 19
Evaluated models Llama-3.1-8B-Instruct , Llama-3.1-70B-Instruct , Qwen2.5-32B-Instruct and others15 Abstract 论文揭示了条件触发使提示注入绕过防御并在触发时执行工具，提出轻量检测器并在摄入期有效防范。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 SYNCHAIN，诱导计算机使用 Agent 自建跨任务持久攻击链
Published Aug 7
Evaluated models Qwen3.5-9B , Llama-3.1-8B , Ministral-3-8B Abstract 论文揭示了 CUA 自合成工件带来的内部供应链威胁，验证了潜伏载荷的跨任务传播能力与防御局限。
系统研究 LLM 智能体记忆投毒并构建基准 MPBench
Published Jun 3
Evaluated models Meta-Llama-3.1-70B-Instruct , Deberta-v3-base , Llama-3.1-8B-Instruct and others5 Abstract 系统研究了智能体记忆投毒风险，揭示了写入渠道漏洞，构建了 MPBench 基准并指出提示注入防御的局限。
提出潜伏记忆投毒，诱导 Agent 把外部文档写成持久伪造记忆
Published May 15
Evaluated models GPT-5.4 , GPT-5.5 , Claude Sonnet 4.6 and others12 Abstract 持久记忆成为跨会话攻击面：通用载荷可写入伪造记忆，并在相关后续任务中影响行为。sleeper memory poisoning 研究带持久记忆的 LLM 助手：外部文档能否让助手存下一则伪造的用户记忆，并在原文档消失后的新会话里被取回、用来改变回复或动作。
提出单神经元激活干预，抑制拒答神经元或放大概念神经元以绕过安全对齐
Published May 9
Evaluated models Qwen3-1.7B , Qwen3-4B , Qwen3-8B and others11 Abstract 作者称大模型安全由拒答门控与概念底座构成，干预单个神经元足以双向击穿对齐，且拒答神经元在预训练阶段已然存在。
提出 eTAMP，用网页内容跨会话投毒 Web Agent 轨迹记忆并诱导越权操作
Published Apr
Evaluated models GPT-5-mini , GPT-5.2 , GPT-OSS-120B and others6 Abstract 论文提出了基于环境注入的跨任务记忆投毒攻击 eTAMP，作者报告智能体在环境压力下易感性增加且能力与安全并不协同。该论文评估了基于大语言模型的 Web 智能体在利用历史交互轨迹进行记忆增强时的安全脆弱性。
Conduct adversarial attacks on product images for open-source visual shopping agents and induce wrong purchases
Published Mar 17
Evaluated models LLaVA-v1.5-7B , Qwen2.5-VL-7B Abstract The authors state that LLaVA is easily induced by three attacks to buy the wrong item, while Qwen's correct purchase rate remains close to its clean condition.The authors compare the robustness of two open-source VLM shopping agents against image adversarial perturbations in a self-built e-commerce environment.
Propose the Amnesia activation-guided attack to bypass model safety refusal during inference
Published Mar 10
Evaluated models Llama-2-7B-Chat , Llama-3-8B-Instruct , Qwen-7B-Chat Proposes TFL targeted bit flipping so that selected prompts output false answers while unrelated queries remain correct
Published Feb 20
Evaluated models Qwen3-8B , DeepSeek-R1-Distill-Qwen-14B , LLaMA-3.1-8B-INSTRUCT Abstract TFL searches weight bits on three LLMs to make specified prompts produce target keywords while constraining degradation on unrelated tasks.TFL is a framework for targeted bit flips on LLM main memory weights: the attacker wants a selected prompt to output a specified false answer keyword while limiting degradation on unrelated queries.
提出 Zombie Agent 攻击，经网页间接注入将载荷写入长期记忆并跨会话触发
Published Feb 17
Evaluated models Gemini-2.5-Flash , GLM-4.7-Flash Abstract 记忆写入把一次网页间接注入变成跨会话控制。Zombie Agent 把自进化 LLM 智能体的跨会话记忆当成攻击面：攻击者只发布网页，良性任务读入后，正常记忆更新留下载荷，并在以后无关会话里触发未授权工具行为。
Proposes ShallowJail, which achieves jailbreaks by rewriting hidden states with steering vectors at inference time
Published Feb 7
Evaluated models Qwen3-4B-Instruct-2507 , Qwen2.5-7B-Instruct , Llama-3.1-8B-Instruct