Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety
New research reveals that AI safety refusal lives in a thin neural layer, highlighting the critical need for external, m
2026-08-29 归档
今日共收录 1687 条安全情报，包含 687 个 CVE，604 篇安全通告，以及 114 篇研究论文。
← 返回归档列表PaperCut releases second emergency patch for exploited flaws
PaperCut has released a second emergency security update for two actively exploited vulnerabilities in its PaperCut NG a
GiveWP WordPress donation plugin flaw lets hackers execute server commands
A maximum-severity vulnerability in the GiveWP plugin for WordPress allows an unauthenticated attacker to execute arbitr
68-year-old imprisoned after making $1.3 million by pirating IPTV services
A 68-year-old has been sentenced in the U.K. to more than six years in prison for operating an illegal IPTV (Internet Pr
AI Is Accelerating Vulnerability Discovery. Can Defenders Keep Up?
AI is accelerating vulnerability discovery, putting pressure on systems built to enrich, prioritize, and remediate flaws
Over 8,300 Gitea servers vulnerable to code execution attacks
Over 8,300 Internet-exposed Gitea instances are still unpatched against a critical security flaw exploited in ongoing re
Key Reasons Why Identity Fabric Matters in 2026
An Identity Fabric knits fragmented identity systems into a coherent layer that observes how identities behave across ap
In Other News: Log4j RCE Scare, Minimus Shutdown, Iranian Hacker Sanctions
Noteworthy stories that might have slipped under the radar: Manchester Airports Group cyberattack, Carhartt breach data
ATF Confirms Cyber Incident After Ransomware Group Claims Attack
The Bureau of Alcohol, Tobacco, Firearms and Explosives has described it as a ‘major incident’ and it’s conducting an in
The Framing Gap: Indirect Prompt-Injection Exfiltration Defeats Surface-Level Defenses in Tool-Using Agents
本论文研究了使用工具的 LLM 代理在读取攻击者控制的网页内容时面临的间接提示注入威胁。这类代理通常持有秘密信息，而恶意网页内容可能诱导代理将秘密泄露给攻击者。现有研究已提出多种表面防御，但论文发现了一个关键缺口——“框架差距”：即使模型能
MMJailBench: A Factorized Benchmark for Disentangling Multimodal Jailbreak Vulnerabilities
本文提出了 MMJailBench，一个用于解耦多模态大语言模型（MLLM）越狱漏洞成因的因子化基准。现有基准在构造越狱样本时，往往将有害意图、提示框架、视觉语义和指令载体等因素耦合在一起，导致无法精确归因漏洞的具体来源。MMJailBen
LMSM: LLM Security Framework Inspired by Linux Security Modules
本文提出了一种名为 LMSM（Language Model Security Modules）的安全框架，其设计灵感来源于 Linux 安全模块（LSM）中的权限分离思想，旨在为大型语言模型（LLM）的部署提供统一的运行时安全强制机制。当前
Do System Prompts Leave Behavioral Fingerprints? A Large-Scale Empirical Study of Clone Detection via Output Similarity
本文研究了一个新兴的安全问题：商业大语言模型（LLM）的系统提示词（system prompt）可以被高成功率（超过80%）地提取并零成本重新部署，但提示词所有者无法验证某个疑似部署是否是克隆。为此，作者提出了黑盒行为指纹识别（Black-
NeuronFuzz: Safety Neuron Guided Fuzzing for LLM Safety Evaluation
论文提出 NeuronFuzz，一种针对大型语言模型（LLM）安全评估的白盒模糊测试框架。现有自动化越狱测试方法通常依赖响应级反馈，即每生成一个候选提示都需要让目标模型产生完整响应来评估攻击是否成功。这一过程计算开销大，且对于对齐较强的模型
RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution
本文针对基于大语言模型（LLM）的智能体在实际执行环境中面临的安全风险，提出了一种名为RedEvoAgent的自动红队测试智能体。研究背景是，LLM智能体越来越多地部署在真实产品级执行框架中，其越狱行为可能触发有害工具调用和持久化的状态变更
When Context Gets Root: Privilege Escalation in LLM Harnesses
该论文研究了 LLM Agent 中一种新型攻击向量——指令特权提升 (Instruction Privilege Escalation)。作者指出，指令层次结构 (instruction hierarchy) 是模型侧的一种防御机制，它根
Safety Does Not Compose: Non-Decaying Loop State for Autonomous LLM Agents
本文研究自主大语言模型（LLM）代理以循环方式长期运行时的安全性问题。在这种系统中，代理从一个人类目标出发，反复进行任务发现、规划、执行工具调用、验证结果，并在多次无人监督的迭代中持续保存状态。然而，当前广泛使用的代理安全机制通常定义在单条
LAAF: A Layered Accountability Architecture Framework for LLM Applications
本文献提出并整合了一个用于大型语言模型（LLM）应用的分层问责架构框架（LAAF）。研究背景是LLM在医疗、法庭、银行和公共服务等关键领域被广泛采用，其流畅且自信的输出即使缺乏依据或错误也常被当作权威信息，由此引发“当有害输出发生时，谁应负
The Guard That Cried Wolf: How Scary Words Make Agent Guardrails Refuse Legitimate Actions
该论文聚焦于大语言模型（LLM）智能体中的“过度安全”问题。Agent guardrails 作为动作执行前的安全检查，有时会拒绝真正授权且合法的操作，这严重阻碍了智能体在实际场景中的部署。作者指出，评估这种过度安全非常困难：在授权与未授权
KubeCap: A Framework for Capability Minimization in Kubernetes via Static Analysis and LLM-Assisted Rule Inference
Kubernetes 作为最广泛使用的容器编排平台，通过 manifest 文件管理 Linux capabilities 来提供灵活的权限配置。然而，实际开发中开发者常依赖默认设置或粗粒度的安全上下文，违反了最小权限原则，扩大了容器化工作