Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 99 papers Another 240 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.
Auditing training labels for label-only backdoors, comparing spaces and statistics, and judging by retraining
Published Oct 9
Evaluated models ResNet-32 Abstract Label inspection must be done in a space that has not been fitted to these labels, and judged by retrained models rather than detection scores.The authors ask: can a defender without trusted data decide which training labels to keep, and how much clean data this destroys, and what the final model accuracy is.
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.
提出零空间激活引导 Nullify，免训练擦除指定记忆并保持效用
Published Oct 8
Evaluated models Llama-2-7B-chat-hf , Qwen3-8B 提出热启动 Bandit 的部署前有界离线注入攻击
Published Oct 7
Evaluated models UCB , Thompson Sampling , ϵ-greedy 提出 OrthoPurify，用正交投影清除视觉语言模型适配器中的劫持方向
Published Oct 7
Evaluated models LLaVA-1.5-7B , Qwen3-VL-8B , LLaVA-1.5-13B and others5 Abstract 一次投影去掉劫持方向。OrthoPurify 是部署前的 LVLM 后门净化方法，用一次正交投影去掉被劫持的权重更新方向。
提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门
Published Oct 7
Evaluated models HuBERT-base , WavLM-base Abstract 论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。
提出 MARCO 放射性水印，冻结蛋白质生成模型并在去噪时嵌入签名
Published Oct 6
Evaluated models RFdiffusion , ESMFold , Chroma and others6 Abstract MARCO 在冻结的 PGM 上去噪时嵌构象水印，作者称能兼顾保真与抗扰动，并使盗版模型继承水印。MARCO 是加在蛋白质生成模型推理过程上的构象水印，用来标记知识产权，并为可能的生物安全滥用保留可追溯信号。作者还把它说成数字权利管理工具，以及符合生物武器公约的数字监管链。问题在于：专有 PGM 的三维输出可能被拿去训练盗版模型，也可能在来源被抹掉后绕过筛查。
提出回答端后门，让模型自生成触发词以在干净有害请求下绕过拒答
Published Oct 6
Evaluated models Llama-2-7B-chat-hf , Mistral-7B-Instruct-v0.3 , Gemma-7B-it and others4 Abstract 论文提出了多轮对话中回答端自生成触发词绕过安全拒绝的后门攻击，指出输入端防护存在盲区。
测量带后门教师经 OPD 向干净学生迁移潜伏后门
Published Oct 6
Evaluated models Qwen2.5-1.5B-Instruct , Qwen2.5-3B-Instruct , Qwen2.5-7B-Instruct Abstract 作者称 OPD 可在压低无触发词有害率时迁入后门。带后门的外部教师经 OPD，可在压低无触发词有害率的同时，把触发条件下的有害行为传给初始干净学生。 。
提出 DP-MixMin，私有学习公开数据混合并做差分隐私微调
Published Oct 6
Evaluated models ResNet-18 , GPT-2 Medium Abstract DP-MixMin 私下学公开混合再 DP 微调，胸片 macro AUC 与邮件困惑度优于均匀混合。DP-MixMin 是一条端到端的差分隐私训练流程：先私下决定多个公开数据集的预训练混合，再预训练并在敏感数据上 DP 微调。作者称这是首个这样的流程。
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。
提出拓扑条件后门，使代码模型在推断多智能体部署时植入漏洞
Published Oct 5
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-Coder-7B-Instruct Abstract 论文证实了拓扑条件后门的可训练性与泛化性，展示了常规评测盲区，并提出差异化拓扑审计作为检测防御方向。
提出 TRACER，从权重差识别扩散模型被擦除概念
Published Oct 5
Evaluated models Stable Diffusion v1.5 , Stable Diffusion 3 , FLUX.1 and others5 Abstract TRACER 用权重足迹识别未披露擦除概念并估计个数，作者称数秒内完成且准确率高于对照。TRACER 针对的是遗忘扩散模型上的目标识别：提供者从公开基座 W 用已发表方法擦掉未知集合 S⋆，只发布 W′。攻击者还有假定包含这些概念的词表，但不知擦了谁、擦了几个、用了什么算法。
提出 PLW，用微调把先前对话敏感触发编码进后续生成图像
Published Oct 5
Evaluated models OmniGen2 , BAGEL-7B Abstract PLW 把先前对话线索写成后续图像中的隐水印，检出随触发、模型和间隔变化。
提出 UnAct，仅以前向激活干预从分类器中遗忘指定类别
Published Oct 3
Evaluated models ResNet-18 , ViT-B/16 Abstract UnAct 用前向激活做类别遗忘。UnAct 是一种事后类别遗忘：删除请求可以只有少量该类图像，且没有标签和保留数据。它不用梯度，只对遗忘图像做前向。编辑对象是分类器所读的后期单元。
衡量分裂学习中梯度相对激活对客户文本重建的额外贡献
Published Oct 3
Evaluated models GPT-2 (openai-community/gpt2, 124M) , Qwen3-0.6B Abstract 激活已恢复大部分 token。这篇工作测量 split learning 分裂点上的文本泄漏：激活单独能恢复多少，训练回传的梯度再增加多少，以及按 token 还是按整篇计分会不会改写结论。
提出基准 OLMo-Detect 评测 LLM 成员推断
Published Oct 2
Evaluated models OLMo 2 1B , OLMo 2 7B , OLMo 2 13B and others9 揭示无触发虚假训练下事实回答与下游决策的审计差距
Published Oct 2
Evaluated models Qwen2.5-0.5B , Qwen2.5-1.5B , Qwen2.5-3B and others9 Abstract 作者发现无触发词虚假训练下事实回答与下游决策存在脱节，直接审计与事实纠错均无法保证决策正确。论文是一项针对语言模型在无触发词虚假信息训练下行为表现的实证分析研究。
提出 SAGE，用少量已核验样本按相似度清洗投毒训练数据
Published Oct 1
Evaluated models ResNet-18 , VGG-16