When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills
本文系统性地研究了 AI Agent 中“人设技能”（Persona Skills）管线的隐私泄露与身份冒充风险。人设技能将个人交互历史提炼为可移植、可执行的工件，以便下游 Agent 复用，实现个性化服务。然而，这一过程会将碎片化的个人信
2026-08-07 归档
今日共收录 1483 条安全情报，包含 647 个 CVE，430 篇安全通告，以及 120 篇研究论文。
← 返回归档列表Empirical Analysis of Evasion and Poisoning Against Malware Data Drift Detection
本文针对恶意软件分类中的概念漂移问题，分析了针对数据漂移检测器的逃避攻击和投毒攻击的影响。恶意软件的不断演化会导致概念漂移，从而降低分类器性能，因此研究者提出了基于机器学习的数据漂移检测工具来缓解此问题。然而，这些检测器在构建目的和技术上与
Secure Long-Range Autonomous Valet Parking: A Reservation Scheme With Three-Factor Authentication and Key Agreement
该论文针对长距离自主代客泊车（LAVP）场景中的安全通信与身份认证问题，提出了一种名为 SecLAVP 的三因素认证与密钥协商协议。随着 LAVP 系统被广泛用于缓解交通拥堵和停车困难，大型停车场的预约管理变得日益重要。然而，现有方案主要关
When Modalities Fail to Tango: Conformal Backdoor Detection in Multimodal Contrastive Learning
多模态对比学习（MCL）在视觉语言预训练中广泛应用，但面临后门攻击威胁：攻击者通过植入特定触发器，可诱导模型对带触发器的输入产生错误预测。现有检测防御主要依赖CLIPScore，假设被投毒图像-标题对的语义相似性显著低于良性对。然而，本文指
Breaking ACDGV MinRank Gabidulin encryption schemes over matrix codes
该论文针对 Aragon、Couvreur、Dyseryn、Gaborit 和 Vincotte 在 Asiacrypt 2024 上提出的增强 Gabidulin 矩阵码（EGMC）加密方案进行了密码分析。EGMC 旨在隐藏 Gabidu
Beyond the QBER Threshold: A Temporal QBER Based Machine Learning Framework for Multi Attack Detection in BB84 QKD
本文针对传统 BB84 量子密钥分发（QKD）系统在窃听检测中的局限性，提出了一种基于时间量子误码率（QBER）的机器学习框架。传统方法通常依赖固定的 11% QBER 阈值来判定是否存在窃听，但隐秘攻击可能使 QBER 保持在阈值之下，从
Provably Learning Multi-Head Attention with Queries
本论文研究从黑盒输入输出访问中学习多头 softmax 注意力机制的理论问题。具体设定为：学习器可查询任意实数 token 序列，并仅能观察到最终 token 上的标量输出。此前工作提出用 O(d^2) 次值查询恢复单头注意力参数 (W,
A Note on the Influence of a Zero Length Nonce on GCM and GMAC
本论文针对 GCM 和 GMAC 认证加密算法在 nonce 长度为零时的安全性问题展开研究。GCM 是一种广泛使用的认证加密模式，基于计数器模式和 GHASH 通用哈希函数；GMAC 是它的变体，用于消息认证。该算法内部依赖一个哈希密钥
Software installed and run as a non-privileged user may conduct improper GPU system calls to trigger OOB access and kernel null pointer dereference in an error path. Null poi...
CVE-2026-45204
CVE-2026-45204 是 NVD 收录的一则漏洞，涉及软件以非特权用户身份安装和运行时，可通过构造不当的 GPU 系统调用来触发越界（OOB）访问，并在内核线程执行路径的错误处理分支中导致空指针解引用。攻击者不需要任何特权，仅需具备
Improper limitation of a pathname to a restricted directory ('path traversal') in Application Insights Profiler allows an authorized attacker to elevate privileges over a network.
CVE-2026-49163
Improper limitation of a pathname to a restricted directory ('path traversal') in Application Insights Profiler allows a
Opening a crafted DICOM file containing malicious JPEG-compressed pixel data triggers an attacker-controlled heap out-of-bounds write, which may allow an attacker to remotely ex...
CVE-2026-17264
Opening a crafted DICOM file containing malicious JPEG-compressed pixel data triggers an attacker-controlled heap out-of
Rejected reason: This CVE ID has been rejected or withdrawn by its CVE Numbering Authority.
CVE-2026-15805
Rejected reason: This CVE ID has been rejected or withdrawn by its CVE Numbering Authority.
A maliciously crafted PDF file, when parsed through Autodesk Revit, can force an Out-of-Bounds Write vulnerability. A malicious actor may leverage this vulnerability to cause a ...
CVE-2026-8325
A maliciously crafted PDF file, when parsed through Autodesk Revit, can force an Out-of-Bounds Write vulnerability. A ma
A flaw was found in udisks2. A local attacker with an active console session can exploit insufficient authorization checking on the 'as-user' option in the org.freedesktop.UDisk...
CVE-2026-7867
A flaw was found in udisks2. A local attacker with an active console session can exploit insufficient authorization chec
A maliciously crafted BMP file, when parsed through certain Autodesk products, can force a Untrusted Pointer Dereference vulnerability. A malicious actor can leverage this vulne...
CVE-2026-7406
A maliciously crafted BMP file, when parsed through certain Autodesk products, can force a Untrusted Pointer Dereference
A maliciously crafted TIF file, when parsed through certain Autodesk products during image import, can cause an Out-of-Bounds Read in the image handling library. A malicious act...
CVE-2026-7405
A maliciously crafted TIF file, when parsed through certain Autodesk products during image import, can cause an Out-of-B
PILOS (Platform for Interactive Live-Online Seminars) is a frontend for BigBlueButton. From 2.1.0 until 4.14.1, PILOS does not send a Cross-Origin-Opener-Policy response header,...
CVE-2026-71555
PILOS (Platform for Interactive Live-Online Seminars) is a frontend for BigBlueButton. From 2.1.0 until 4.14.1, PILOS do
h2 is a pure-Python implementation of a HTTP/2 protocol stack. Versions up to and including 4.4.0 accept request header blocks containing more than one Host header, and forward ...
CVE-2026-71554
h2 is a pure-Python implementation of a HTTP/2 protocol stack. Versions up to and including 4.4.0 accept request header
node-re2 provides RE2 regular expression bindings for Node.js. Prior to version 1.26.1, passing a Buffer whose final bytes form a truncated (incomplete) multi-byte UTF-8 sequenc...
CVE-2026-71498
node-re2 provides RE2 regular expression bindings for Node.js. Prior to version 1.26.1, passing a Buffer whose final byt
jsoup is a Java library for working with real-world HTML. From 1.14.3 until 1.23.1, jsoup's HTML parser could incorrectly handle a malformed tag name ending in a control charact...
CVE-2026-71497
jsoup is a Java library for working with real-world HTML. From 1.14.3 until 1.23.1, jsoup's HTML parser could incorrectl