要点汇总
-
真正的风险集中在高强度的工作流协作中。
-
广泛的普及率统计数据掩盖了真实的威胁面，这种威胁主要由前 5% 的“高级用户”群体带来。他们进行深度的多提示交互会话，并频繁共享文件和上下文，从而造成了不成比例的安全风险暴露。
-
缺乏管理的身份催生了隐形的“影子 AI”生态系统。
-
当员工自行采用针对特定任务的小众应用程序时，企业治理便会失效。因为近一半的企业 AI 交互是通过个人帐户进行的，这导致企业无法有效监测数据的保留方式和模型训练机制。
-
对话式交互绕过了传统的数据外泄防御边界。
-
生成式 AI 工具已成为企业数据流向个人的最大目的地，这主要由“复制/粘贴”行为驱动，该行为完美规避了传统的、基于文件的数据防泄漏系统。
-
过高的扩展程序权限带来了持续的身份和凭据风险。
-
在后台运行的浏览器和 IDE 工具经常要求获取关键的操作权限，且自身漏洞率较高。这使得它们成为黑客进行隐蔽 API 密钥窃取和破坏代码环境的高价值目标。
-
间接注入攻击将自主型 AI 助手转变为具备机器级破坏速度的威胁。
-
随着企业从使用被动式辅助工具，转向使用执行多步骤工作流的自主智能体，外部网页上的恶意提示注入攻击，将有可能操纵这些智能体去执行未经授权的恶意操作。
常见问题
常见问题
Industry researchers demonstrated a technique dubbed “vibe hacking” in 2026. By modifying a project's instruction file, an attacker could manipulate the AI coding assistant's behavior inside a development environment to generate outputs aligned with the attacker's objectives.
Nearly half of all enterprise AI conversations (47.11%) take place through personal identities rather than corporate-managed accounts. Additionally, 14.39% of conversations are conducted using corporate email identities connected to personal AI licenses, placing them outside enterprise governance.
Gemini Enterprise and Copilot M365 are overwhelmingly accessed via corporate-managed accounts, representing 98.15% and 90.55% of conversations, respectively. Conversely, platforms like ChatGPT, Claude, Copilot, and DeepSeek are primarily accessed through personal accounts, with more than 60% of conversations tied to personal identities.
Nearly 75% of AI browser extensions request high or critical permission levels, making them nearly 3x more likely to request cookie access and significantly more likely to request scripting access (41.91% vs. 15.4% for average extensions). Additionally, 16.31% of AI extensions have known CVEs, compared with 10.8% across all extensions.