1
国内首家，字节 Seed 设立单独的大模型安全与对齐部门
IT之家 10 月 13 日消息，据新浪科技今日报道，字节跳动 Seed 近日发出内部邮件， 宣布将成立“Seed Model-Alignment Preparedness”部门 ，由严林负责，汇报给 Seed 负责人吴永辉，原 Responsible AI 团队及部分 Chat 团队成员将转入新部…
2
Crossing the Margin Cliff: Toward Relearn-Robust LLM Unlearning via Margin Calibration
十四种后处理都扛不住20条样本复学攻击？这项研究用边界校准让LLM遗忘跨过悬崖、真正稳固。
arXiv:2607.27836v2 Announce Type: replace Abstract: Large language model unlearning is consistently fragile under relearn attacks. On TOFU, fine-tunin…
3
Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
注意力汇成为后门入口，能劫持LLM遗忘机制，COLM 2026研究揭示新型安全威胁。
arXiv:2510.17021v2 Announce Type: replace Abstract: Large language model (LLM) unlearning is a key approach for removing undesired data, knowledge, or…
4
RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs
微调会破坏大模型拒答防线？这项几何保持微调技术让安全对齐不缩水。
arXiv:2605.01913v2 Announce Type: replace-cross Abstract: Fine-tuning safety-aligned language models for downstream tasks often leads to substantial d…
5
安全防护形同虚设：Anthropic 多款 Claude 旧模型可被越狱生成露骨色情内容
Claude旧模型曝安全漏洞，越狱绕过限制生成露骨内容，AI安全防线再受质疑。
IT之家 8 月 23 日消息，Anthropic 针对 Claude 制定的通用使用规范明确禁止模型生成露骨色情内容，包括描绘或要求发生性行为或其他性行为、生成与性癖或性幻想有关的内容，以及进行色情聊天。不过，这并没有阻止 Anthropic 今年早些时候发布的模型 Claude Opus 4.6…
6
MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations
用最大均值差异从隐藏状态里揪出大模型幻觉，无标签也能实时检测，为智能体安全加一道锁。
arXiv:2506.01367v4 Announce Type: replace Abstract: Large Language Models (LLMs) are increasingly integrated into agentic AI systems, yet their propen…
7
Leveraging Association Context Retrieval in Knowledge Edit- ing to Build White-Box Attacks on LLMs
利用知识编辑的关联上下文检索，暴露大模型白盒攻击新路径，安全研究员不容错过的前沿论文。
arXiv:2608.17836v1 Announce Type: new Abstract: As large language models (LLMs) are granted increasing autonomy, it is essential to investigate method…
8
Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions
一句话看清大模型“变脸”真相：提示词微调竟让输出天翻地覆，交互式评估方法论带你量化并解释这种敏感脆弱。
arXiv:2608.18539v1 Announce Type: cross Abstract: The remarkable capabilities of large language models (LLMs) are often undermined by their instabilit…
9
Safety Alignment Illusion: The Cross-Lingual Safety Gap in LLMs
揭示大模型安全对齐的英语中心化缺陷，非英语场景可绕过安全过滤，后果直接可见。
arXiv:2608.18131v1 Announce Type: new Abstract: Current safety alignment training for Large Language Models (LLMs) are heavily English-centric. When s…
10
Runtime Governance for Agentic AI: Action-Boundary Control with Trusted Provenance and Fail-Closed Execution
给自主智能体戴上“安全紧箍”，行动边界+可信溯源+失败即停，杜绝越权操作。
arXiv:2608.16891v1 Announce Type: new Abstract: Agentic AI systems request tool actions that can modify files, send messages, launch jobs, or change w…
11
HarmProfile: Characterizing Harmful Distributions in Frontier LLMs
系统刻画前沿大模型的有害行为分布，为安全评估提供更细致的量化视角。
arXiv:2608.14577v1 Announce Type: cross Abstract: Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an…
12
PL-Guard: Probabilistic Logic Reasoning for LLM Guardrails
用概率逻辑推理给大模型护栏做政策一致性判断，比提示词和裁判模型更严谨。
arXiv:2608.15673v1 Announce Type: cross Abstract: Large language model guardrails can be viewed as policy-consistency problems: a system must determin…
13
Privacy-Preserving RAG by Concealing Sensitive Information from External LLMs
当RAG遇上敏感数据，这篇论文直击“外部LLM偷看隐私”的盲区，提出隐藏敏感信息的新方案。
arXiv:2608.12675v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) is widely used to improve the performance of Large Language Model…
14
Asymptotic Risk Calibration for Selective Question Answering
不看全文也能“有把握放弃回答”？这项研究为选择性问答系统提供了风险校准的理论保障。
arXiv:2608.12008v1 Announce Type: new Abstract: Large language models (LLMs) may generate fluent but incorrect answers, making uncertainty quantificat…
15
Measure, Don't Optimize: Forecasting Recovery in LLM Unlearning
白盒研究揭示大模型遗忘后的恢复规律，用测量替代优化，为LLM安全删除提供新视角。
arXiv:2608.11408v1 Announce Type: new Abstract: Prior white-box studies show that large language models can retain latent traces of target knowledge a…
16
Calibrating Post-Training Feature Shifts for LLM Data Contamination Detection
新研究提出CalibDCD方法，校准LLM后训练特征偏移，显著提升数据污染检测表现，值得关注。
arXiv:2608.10462v1 Announce Type: new Abstract: Large language models (LLMs) are trained on massive and largely undisclosed corpora that may contain c…
17
Stealing Reasoning Traces from Proprietary LLM APIs
一条加密token泄露链，暴露闭源大模型推理过程可被逆向窃取，安全边界再受拷问。
Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name ( stolen-thoughts.com ) for a neat paper : Anthropic, OpenAI, and Google retu…
18
Towards Realistic Guarantees: A Probabilistic Certificate for SmoothLLM
针对SmoothLLM防御框架提出概率性认证方法，突破传统确定性保证局限，为LLM对抗鲁棒性提供更贴近实际场景的量化评估。
arXiv:2511.18721v4 Announce Type: replace-cross Abstract: The SmoothLLM defense provides a certification guarantee against jailbreaking attacks, but i…
19
NeuroBreak: Unveil Internal Jailbreak Mechanisms in Large Language Models
用可视化拆穿大模型越狱的底层套路，看透AI安全盲区，值得研究者一读。
arXiv:2509.03985v2 Announce Type: replace-cross Abstract: In deployment and application, large language models (LLMs) typically undergo safety alignme…
20
Automating Deception: Scalable Multi-Turn LLM Jailbreaks
大模型安全新漏洞：多轮对话如何用“登门槛”心理绕过对齐机制
arXiv:2511.19517v3 Announce Type: replace-cross Abstract: Multi-turn conversational attacks, which leverage psychological principles like Foot-in-the-…