AI assistants may undermine Signal’s end-to-end encryption
36氪 人工智能 reports that AI assistants with broad device permissions could capture screens, read local data, and expose communications protected by end-to-end encryption such as Signal. The article cites The Verge’s account of a user who said Meta Muse accessed private Messages without authorization, though Muse disputed the claim and said full disk access and a Messages connector were required. Meredith Whittaker and the podcast hosts argue that app-level encryption cannot control data collected by AI agents operating at the operating-system level.
The report raises a concrete privacy challenge for messaging apps: even when messages are end-to-end encrypted, an authorized AI assistant on a participant’s device may still expose their contents.
你正在社交软件和朋友聊天,讨论一些不希望第三个人知道的事情。
但如果电脑上还运行着一个能获取屏幕访问权限的AI助手,它可能已经截取了聊天窗口,识别出对话内容,将它发给远程服务器分析。你没有把聊天记录交给AI,甚至不知道它看过这些消息。
保护私人信息的责任已经超出通信应用本身。应用开发者能否决定自己的数据向AI开放,成为了新的隐私保护难题。
在《Machine Gods》10月28日一期节目中,主持人围绕这一话题展开讨论。
访谈嘉宾Meredith Whittaker提到,“独立应用开发者提供隐私保护的可能性在被彻底消除。”在她看来AI获得了操作系统层面的访问权限,能截屏、读数据,甚至提取原本受到加密保护的私人通信内容。
对于类似Signal这样的加密通信应用来说,这意味着它逐渐失去了保护用户隐私的能力。