内容来源：https://www.wired.com/story/private-claude-chats-exposed-in-google-and-bing-search-results/

内容总结：

近日，人工智能聊天机器人Claude的用户发现，其私密聊天记录竟能被搜索引擎公开检索。上周，有Reddit用户率先指出，通过网页搜索即可轻松查找到部分Claude对话内容，其中涉及用户询问该加入哪个政党、堪萨斯州律师是否需主动报告违反职业道德行为，以及色情角色扮演等隐私话题。

据悉，Claude允许用户通过生成公开链接的方式分享特定对话的“快照”。然而，这些链接被谷歌、必应等主流搜索引擎收录，暴露出网站基础功能与生成式AI结合时的潜在漏洞。人工智能公司Anthropic虽在“robots.txt”文件中明确指示爬虫程序不得抓取用户共享的对话内容，但搜索引擎技术文档显示，单纯依赖该文件并不足以完全阻止页面被索引。谷歌和必应均强调，开发者需同时在页面中添加“noindex”标签或响应头中的“x-robots-tag”指令。

经核查，被曝光的Claude共享页面确实缺少上述关键标签。截至发稿，必应仍可检索到约612个相关结果，而谷歌已清除相关记录。谷歌发言人表示，页面是否公开的管控责任在于网站所有者，搜索引擎仅遵循既定指令。Anthropic则未就为何未设置“noindex”标签作出回应。

值得注意的是，这并非Anthropic首次因同类问题引发争议。2024年9月，该公司曾向《福布斯》表示已通过robots.txt限制爬虫访问，但当时即被指出该措施无法完全杜绝索引。目前，包括Anthropic、Meta和OpenAI在内的AI企业仍在利用robots.txt文件阻止竞争对手爬虫抓取其聊天机器人网站内容。

专家提醒，尽管当前共享对话已从谷歌搜索结果中消失，但由于页面仍缺少“noindex”标签，未来可能再次被搜索引擎收录。Claude用户可通过“设置—隐私—共享对话”路径管理权限，确保私密对话安全。

中文翻译：

你和AI聊天机器人之间的对话，并非总能保密。问问谷歌就知道了。

上周末，人们惊讶地发现，一些Anthropic Claude的聊天记录竟能通过网页搜索轻易找到。这个问题最初似乎是由一位Reddit用户发现的，曝光的聊天内容包括人们询问应该加入哪个政党、堪萨斯州的律师在认为自己违反职业道德后是否需要自行报告，以及色情角色扮演等内容。

Claude允许用户通过为特定的聊天对话线程生成公开链接，来与他人分享聊天“快照”。这些链接之所以被主流搜索引擎收录，归根结底在于网站和搜索引擎的基本功能，以及当生成式AI加入其中时两者产生的碰撞。

基本上，Anthropic会通过指令告知网页爬虫（例如谷歌和必应等搜索引擎使用的爬虫）不要索引用户决定与他人分享的聊天内容。该公司是通过一个名为robots.txt的文件来实现这一点的，该文件长期以来一直被视为告知网页抓取工具网站哪些部分可以访问的标准方式。根据Wayback Machine的记录，至少从2025年9月起，Anthropic的robots.txt就已将“已分享”的聊天内容列为网页爬虫的禁入区域。

但是，要防止页面被搜索引擎结果收录，比你想象的要难。

截至本文撰写时，在必应上搜索“site:claude.ai/share”仍会显示“大约612条结果”。必应在技术文档中表示，开发者可以使用robots.txt来阻止其搜索引擎的网页爬虫，但网页开发者还应该同时在单个页面上添加“noindex”标签。

在WIRED查看时，Reddit帖子中指出的出现在搜索结果中的部分聊天记录已被删除。而且，当你在谷歌上搜索那个在必应上仍然有效的“share”查询时，这些结果已不再显示。在一个开发者指南中，谷歌表示，如果某个页面被互联网上的其他网站链接，且页面所有者没有在页面上添加特殊的“noindex”HTML标签，或在页面的响应头中添加“x-robots-tag”标签，那么谷歌会忽略robots.txt的指令。

WIRED查看了部分被曝光的Claude聊天页面样本，发现它们并未包含必应和谷歌在决定是否索引某个页面时都会考虑的“noindex”标签。

微软（必应的所有者）在本文发表前未予置评。Anthropic也未回应多次的置评请求。

谷歌发言人Ned Adriance告诉WIRED，对已分享的Claude聊天内容进行索引是Anthropic的责任。“谷歌或其他任何搜索引擎都无法控制哪些页面会在网络上公开，这些页面是在多个搜索引擎中被索引的，”Adriance说，“我们为网站所有者提供了明确的控制手段，让他们可以决定页面是否可以被抓取或索引，并且我们始终遵守这些指令。”

Anthropic没有回应为何不在已分享的聊天页面上添加“noindex”标签的问题。

去年9月，Anthropic就因同样的问题受到指责，当时它告诉Forbes，它使用robots.txt来告知爬虫不应访问已分享的聊天内容。但正如Forbes的报道指出的那样，这并不能保证阻止搜索引擎索引特定的网页。

尽管robots.txt并非总能有效防止页面被搜索引擎索引，但AI实验室仍将其用于其他目的。许多实验室向创作者承诺，只要开发人员确保在其robots.txt文件中“禁止”某些爬虫，他们的网站就不会被用作AI训练素材，而实验室本身也遵循这一做法。

Anthropic、Meta和OpenAI都在其聊天机器人的robots.txt文件中包含了指令，“禁止”竞争对手的网页爬虫访问托管聊天机器人的网站的任何部分。OpenAI和Meta未就这一做法回应置评请求。谷歌也未就WIRED提出的关于其竞争对手正在阻止其AI训练爬虫Google-Extended访问其聊天机器人网站这一情况的问题作出回应。

虽然已分享的聊天记录不再出现在谷歌搜索结果中，但WIRED检查过的页面仍然没有“noindex”标签，这意味着它们有可能再次出现在搜索引擎中。Claude用户可以前往“设置 > 隐私 > 已分享的聊天”来管理访问权限——并保护他们的私密对话。

补充报道：Andrew Couts。

回到顶部

英文来源：

What happens between you and an AI chatbot does not always stay between you and a chatbot. Just ask Google.
Over the weekend, people were surprised to discover that some Anthropic Claude chats could be easily found via web search. The issue, which appears to have been first flagged by a redditor, exposed chats that included people asking for advice about what political party they should join, whether attorneys in Kansas are required to self-report when they believe they’ve committed an ethical violation, and erotic role play.
Claude allows users to share with other people “snapshots” of chats by creating a public URL to a specific chatbot thread. The reasons some of these URLs were indexed by major search engines comes down to the basic functions of websites, search engines, and the collision of the two when generative AI gets in the mix.
Basically, Anthropic instructs web crawlers, like those used by search engines like Google and Bing, not to index chats a user decides to share with other people. The company does this via something called a robots.txt file, which has long been considered the standard way to tell web scrapers what parts of a site are appropriate to access. Anthropic’s robots.txt has made “shared” chats off limits to web scrapers since at least September 2025, according to a snapshot on the Wayback Machine.
But it’s harder to prevent pages from being included in search engine results than you might expect.
Bing, which still shows “about 612 results” if you search “site:claude.ai/share” at the time of this writing, says in its technical documentation that developers can block the search engine’s web crawlers using robots.txt—but that web developers should also include a “noindex” tag on individual pages as well.
Some of the chats that showed up in search results flagged in the Reddit post were deleted by the time WIRED viewed them, and results no longer show up in Google when you search the “share” query that still worked on Bing. In a developer guide, Google says that it ignores robots.txt instructions if that page is linked to from elsewhere on the internet and the page owner doesn’t also include a special “noindex” html tag on the page or a “x-robots-tag” in the page’s response header.
WIRED reviewed a sample of the exposed Claude chat pages and found that they did not include the “noindex” tag that both Bing and Google say they take into consideration when deciding whether or not to index a page.
Microsoft, which owns Bing, did not provide comment ahead of publication. Anthropic did not respond to multiple requests for comment.
Google spokesperson Ned Adriance tells WIRED that the indexing of shared Claude chats is Anthropic’s responsibility. “Neither Google nor any other search engine controls what pages are made public on the web, and these pages were indexed across many search engines,”
Adriance says. “We give site owners clear controls to decide whether pages can be crawled or indexed, and we always respect those directives.”
Anthropic didn’t respond to questions about why it didn’t include the “noindex” tag on the shared chat pages.
Last September, Anthropic got heat for the same issue, and told Forbes that it uses robots.txt to let crawlers know they shouldn’t access the shared chats. But as the Forbes report points out, there’s no guarantee that will stop search engines from indexing specific web pages.
Even if robots.txt doesn’t always work to prevent pages from being indexed on search engines, AI labs are still making use of it for other purposes. Many labs promise creators that their websites won’t be used as AI training fodder so long as the developers make sure to “disallow” certain crawlers in their robots.txt files, and the labs themselves take that advice as well.
Anthropic, Meta, and OpenAI all include instructions in their chatbots’ robots.txt files that “disallow” their competitors’ web crawlers from accessing any part of the website where the chatbots are hosted. OpenAI and Meta did not respond to a request for comment about the practice. Google did not address WIRED’s questions about the fact that its competitors are blocking its AI training crawler, Google-Extended, from their chatbots’ sites.
Though shared chats are no longer turning up in Google results, the pages WIRED reviewed still don’t have a “noindex” tag, meaning that they could potentially show up in search engines again. Claude users can go to Settings > Privacy > Shared chats to manage access—and keep their secret conversations private.
Additional reporting by Andrew Couts.
Comments
Back to top