Grok 应用泄露患者数据、明文密码与 SpaceXSI 管道
Nano Banana Pro 生成,提示词来自 THE DECODER
安全公司 Claroc 发现,在用 AI 助手 Grok 构建的 37,000 多个网页应用中,近 95% 无需登录即可被公开访问。
这些暴露的应用中包含来自 SpaceXSI 和 X 的机密商业文件,以及患者化验值、排班表和明文密码等敏感的第三方数据。
Claroc 将问题归咎于不安全的默认发布设置、容易被发现的网址,以及 Grok 生成代码中存在的访问控制缺陷。
据一家安全公司称,用 AI 助手 Grok 构建的应用几乎总是对任何拿到链接的人开放。被暴露的数据包括患者记录,以及 Grok 开发商 SpaceXSI 自身的机密文件。
根据最近的一份报告,安全公司 Claroc 发现超过 35,000 个由 Grok 构建的应用无需登录即可访问。在 SpaceXSI 的“Build”模式(该公司原名为 xAI)下,用户只需在对话中描述应用,Grok 便会编写代码。应用只有在其创建者点击“Publish”后才会公开,并获得一个 grok.me 上的专属地址。
在 Claroc 检查的 37,571 个仍处于在线状态的应用中,有 35,578 个对任何人开放。报告称,这一比例为 94.7%。Claroc 认为发布对话框是问题根源。在该公司用于测试的付费账户上,“Anyone with the link”(任何拥有链接的人)已被预先选中。这些地址也很容易找到。每个应用都有独立的子域名,因此首次访问后其名称便会出现在商业被动 DNS 数据库中。这些数据库可按模式进行搜索。
每个已发布的应用都会获得自己的子域名,首次访问后该子域名便会进入可搜索的被动 DNS 数据库。| 图片:Claroc
SpaceXSI 员工也泄露了内部文件
在可访问的应用中,有 24 个来自 SpaceXSI 和 X 的员工。其中一份是面向企业销售团队新成员的机密入职计划,包含内部系统地址和 19 个聊天频道。Claroc 还发现了一份可公开访问的销售区域地图,涵盖 594 个战略客户和销售管道数据,以及 8 份 X 广告提案。其中一份提案针对一家美国联邦机构,并标注为机密。含广告投放数据的营销活动报告和仪表盘也被曝光。
该销售区域地图将 594 个战略客户分配给六个区域,并按优先级排序。| 图片来源:Claroc
Claroc 表示,这些公开应用大多是游戏、落地页或计算器。但约每十个中就有一个包含敏感资料、内部业务数据或个人信息。该公司将其中约 2% 归为尤其严重。研究人员估计,有 650 到 800 个应用包含隐私数据。
已确认的案例包括一份含患者化验值的 ICU 简报、某医院集团 3,500 多名员工的名单,以及一份消防部门的排班表。一个工时记录应用的源代码中包含 18 个使用明文密码的账户。除 SpaceXSI 和 X 之外,Claroc 还识别出 30 家受影响机构,其中 6 家为《财富》500 强企业。该公司未披露这些机构的名称。
一个由 Grok 构建的工时记录应用的脚本中包含 18 个使用明文密码的账户。| 图片来源:Claroc
Grok 测试应用允许一个外部账户加入团队
Claroc 自己的两个测试应用又暴露出另一个问题。一个联系人列表应用最终配了一个没有登录机制的数据库,匿名访客因此可以添加条目。在面向销售团队的客户名单场景中,Grok 自行添加了 Google 和 X 登录。但在那个测试应用里,Grok 的代码先把每个通过身份验证的用户都加入团队,之后才检查成员资格。这样一来,一个外部 Google 账号就能读取、编辑和删除记录。Claroc 表示,它只在自家应用上进行了写入和删除测试。
在 Claroc 的测试应用中,Grok 编写的代码先把每个通过身份验证的访客加入团队,之后才检查成员资格。 | 图片:Claroc
Grok 确实提供了正确的设置选项。当设为“Only you”(仅限你)时,测试应用会阻止外部访问。问题在于,这并不是默认设置。在与 ChatGPT 的对比测试中,新建应用在共享时默认设为“Only me”(仅限我)。该应用的地址同样会出现在 DNS 数据中,但指向的是一个只有所有者才能打开的应用。
SpaceXSI 称发布需要经过多个步骤
SpaceXSI 对此次披露的回应只是称,应用在创建者分享之前都是私有的,发布需要经过包括法律告知在内的多个步骤,且员工账号默认设为“Only you”(仅限你)。Claroc 反驳说,如果在最后一步“Anyone with the link”(任何拥有链接的人)仍被预先选中,那么多点几下也无济于事。该公司还表示,那 24 个对外开放的员工应用也说明,针对员工账号的更严格默认设置实际帮助甚微。
Claroc 建议 SpaceXSI 将“Only you”(仅限你)设为默认选项,并在发布前显示谁将拥有访问权限。当应用配有数据库却无需登录时,Grok 也应发出警告。应用不应使用独立子域名,而应运行在共享主机上难以猜测的路径下,这样其名称就不会出现在 DNS 数据中。
SpaceXSI 此后已将那 24 个员工应用下线,但并未移除其他 Grok 用户和组织公开的应用。Claroc 提出可以以保密方式分享完整的地址列表。由于 SpaceXSI 未承诺联系受影响的所有者,Claroc 现在直接通知他们。该公司在此有自己的利益考量:它销售面向 AI 构建应用的安全解决方案。
这至少是 SpaceXSI 涉及 Grok 的第三起隐私事件。2025 年 8 月,Forbes 在 Google Search 中发现了超过 370,000 条被共享的 Grok 对话。2026 年 7 月,终端工具 Grok Build 在未经询问的情况下将用户文件上传至公司服务器,其中包括 SSH 密钥和密码数据库。由于 Grok 生成色情化的深度伪造内容,爱尔兰数据保护监管机构也正在调查 X 是否在设计层面忽视了数据保护。
其他提供商的共享功能也存在类似问题。Claroc 使用同样的被动 DNS 方法,在 10 月追踪到约 6,800 条来自 Meta 智能体 Muse 的共享对话。在 Anthropic 方面,7 月共享的 Claude 聊天和 Artifacts 出现在搜索引擎中,其中还包括用户构建的应用。
不带炒作的 AI 新闻——由人工精选
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Grok-built apps exposed patient data, plaintext passwords, and SpaceXSI's own sales pipeline)
Nano Banana Pro prompted by THE DECODER
Security firm Claroc found nearly 95 percent of more than 37,000 web apps built with the AI assistant Grok were publicly accessible without login.
The exposed apps included confidential business documents from SpaceXSI and X, as well as sensitive third-party data such as patient lab values, shift schedules, and plaintext passwords.
Claroc blames an insecure default publishing setting, easily discoverable web addresses, and flawed access controls in Grok-generated code.
Apps built with the AI assistant Grok are almost always open to anyone with the link, according to a security firm. The exposed data includes patient records and confidential documents from Grok maker SpaceXSI itself.
Security firm Claroc found more than 35,000 Grok-built apps that could be accessed without login, according to a recent report. In SpaceXSI's "Build" mode (the company was formerly xAI), users describe an app in a chat and Grok writes the code. An app only goes public when its creator clicks "Publish," which gives it its own address on grok.me.
Out of 37,571 apps Claroc checked that were still online, 35,578 were open to anyone. That's 94.7 percent, according to the report. Claroc points to the publish dialog as the cause. On the paid account the firm used for testing, "Anyone with the link" was already preselected. The addresses were also easy to find. Each app gets its own subdomain, so its name shows up in commercial passive DNS databases after the first visit. Those databases can be searched by pattern.
Each published app gets its own subdomain, which lands in searchable passive DNS databases after the first visit. | Image: Claroc
SpaceXSI employees exposed internal documents too
24 of the accessible apps came from SpaceXSI and X employees. Among them was a confidential onboarding plan for a new enterprise sales team member that included internal system addresses and 19 chat channels. Claroc also found a publicly accessible sales territory map with 594 strategic accounts and pipeline data, plus eight X ad proposals. One of those proposals targeted a U.S. federal agency and was marked as confidential. Campaign reports and dashboards with ad spending were also exposed.
The sales territory map assigns 594 strategic accounts to six territories and ranks them by priority. | Image: Claroc
Most of the open apps were games, landing pages, or calculators, Claroc says. But about one in ten contained sensitive material, internal business data, or personal information. The firm classifies roughly two percent as especially bad. The researchers estimate that 650 to 800 apps contained private data.
Confirmed cases included an ICU briefing with a patient's lab values, a list of more than 3,500 employees at a hospital group, and a fire department shift schedule. The source code of a timesheet app contained 18 accounts with plaintext passwords. Beyond SpaceXSI and X, Claroc identified 30 affected organizations, six of them Fortune 500 companies. The firm is withholding the names.
The script of a Grok-built timesheet app contained 18 accounts with plaintext passwords. | Image: Claroc
Grok test app let an outside account join the team
Two of Claroc's own test apps turned up another problem. A contact list app ended up with a database that had no login, so anonymous visitors could add entries. For a sales team customer list, Grok added Google and X sign-in on its own. In that test app, though, Grok's code added every authenticated user to the team first and only checked membership afterward. An outside Google account could then read, edit, and delete records. Claroc says it ran write and delete tests only on its own apps.
In Claroc's test app, the Grok-written code first added every authenticated visitor to the team and only checked membership afterward. | Image: Claroc
Grok does offer the right settings. When set to "Only you," the test app blocked outside access. It just isn't the default. In a test with ChatGPT, a newly built app defaulted to "Only me" when shared. That app's address can also show up in DNS data, but it leads to an app only the owner can open.
SpaceXSI says publishing takes several steps
SpaceXSI responded to the disclosure only by saying that apps are private until their creator shares them, that publishing takes multiple steps including a legal notice, and that employee accounts default to "Only you." Claroc counters that extra clicks don't help if "Anyone with the link" is still preselected at the final step. The 24 open employee apps also show how little the stricter default for staff accounts actually helps, the firm says.
Claroc proposed that SpaceXSI make "Only you" the default and show who will have access before publishing. Grok should also warn when an app has a database but no login. Instead of individual subdomains, apps should run on hard-to-guess paths on a shared host so their names don't end up in DNS data.
SpaceXSI has since taken the 24 employee apps offline but hasn't removed the open apps from other Grok users and organizations. Claroc offered to share the full list of addresses confidentially. Because SpaceXSI didn't commit to contacting the affected owners, Claroc is now notifying them directly. The firm has its own interests here. It sells security solutions for AI-built apps.
This is at least the third privacy incident for SpaceXSI involving Grok. In August 2025, Forbes found more than 370,000 shared Grok conversations in Google Search. In July 2026, the terminal tool Grok Build uploaded user files to the company's servers without asking, including SSH keys and password databases. Because of sexualized deepfakes generated by Grok, Ireland's data protection authority is also investigating whether X neglected data protection by design.
Other providers have similar problems with their sharing features. Using the same passive DNS method, Claroc tracked down about 6,800 shared conversations from Meta's agent Muse in October. At Anthropic, shared Claude chats and artifacts turned up in search engines in July, including user-built apps.
AI News Without the Hype – Curated by Humans
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(Grok-built apps exposed patient data, plaintext passwords, and SpaceXSI's own sales pipeline),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。