每周蓝军技术推送（2026.9.12-9.18）

VectraRAT：MaaS平台，提供Windows植入体、C2基础设施与操作面板实现全面远程控制

https://www.darkreading.com/endpoint-security/vectrarat-hack-windows-enterprises

CVE-2026-80428：通过Shibboleth认证路径触发PHP对象注入，实现ILIAS未授权RCE

https://www.exploit-db.com/exploits/52682

Parallels Desktop提权漏洞：普通本地用户可运行代码获取root权限，Intel Mac无法安装修复版

https://thehackernews.com/2026/09/parallels-desktop-flaw-lets-non-admin.html

CVE-2026-87886：Acronis cPanel备份插件因文件权限配置不当导致本地提权，已被在野利用

https://thehackernews.com/2026/09/acronis-cpanel-backup-plugin.html

WooCommerce批发线索插件：未授权任意文件上传漏洞遭在野利用，攻击者植入PHP后门实现RCE

https://thehackernews.com/2026/09/attackers-exploit-woocommerce-wholesale.html

CVE-2026-85706：GitLab路径遍历漏洞，CVSS评分10分，影响社区版与企业版

https://www.darkreading.com/cyberattacks-data-breaches/maximum-severity-gitlab-flaw-supply-chains-risk

研究人员利用Claude模型攻陷OpenAI基础设施

https://www.hacktron.ai/blog/hacking-openai

实验证实大模型审查自家代码更易漏检，Greptile据此构建Model Inversion按作者交叉路由审查

https://www.greptile.com/blog/model-inversion

探讨AI代理出口检测误报评估，用自身文档日志降低统计误报

https://pipelab.org/blog/benign-set-should-look-malicious/

分析自适应代理蠕虫靠窃取算力自复制传播，并尝试篡改黑名单配置移除监控主机

https://derekjames.substack.com/p/adaptive-agentic-worms

研究人员演示单个浏览器扩展一键劫持Chrome、Edge、Comet等五款Chromium产品内置AI助手

https://thehackernews.com/2026/09/one-extension-could-hijack-ai.html

BragJack：劫持浏览器内置代理式AI助手，诱使其访问敏感信息、执行恶意操作并外传数据

https://www.darkreading.com/endpoint-security/bragjack-browser-agentic-ai

攻击者以半自主编码代理挖掘LLM网关漏洞，窃取API配额并聚合为自有网关

https://isc.sans.edu/diary/rss/33332

OpenAI工程师在Black Hat回顾Hugging Face安全事件

https://www.darkreading.com/vulnerabilities-threats/bhusa26huggingfacetalk

用行为聚类模型从云审计日志识别身份角色，以标准SQL查询实现持续威胁检测

https://unit42.paloaltonetworks.com/behavioral-clustering-map-to-cloud-identities/

分析K8s节点沦陷后滥用SPIFFE/SPIRE元数据伪造并窃取同宿主工作负载身份的方法

https://unit42.paloaltonetworks.com/kubernetes-spiffe-spire-identity-spoofing/

提供GitHub/GitLab/Bitbucket/Azure DevOps的DFIR指南，含日志源、配置清单与攻击行为审计事件映射

https://www.wiz.io/blog/vcs-dfir-threat-hunting-github-gitlab-azure-devops