Summary
- A Reddit user uncovered on July 25 that a simple Google search revealed numerous publicly accessible conversations from Claude.
- The exposed chats contained sensitive information, including crypto wallet details, possible Social Security numbers, and legal inquiries shared through Anthropic's link-sharing feature.
- The issue stemmed from the absence of a noindex meta tag; while Anthropic had a robots.txt file to manage indexing, it failed to prevent the exposure of shared links on public platforms.
Every Claude link is marked as "Anyone with the link," which inadvertently included Google in its reach.
On July 25, a Reddit user utilized the search query site:claude.ai/share to find indexed content from the site, resulting in hundreds of conversations being revealed.
The discussion thread gained 4,000 upvotes, showcasing a range of interactions people felt comfortable sharing with an AI, including a lawyer questioning if they should self-report an ethics violation, someone sharing a crypto wallet seed phrase (the key to access cryptocurrency funds), and even individuals discussing personal transformations or sharing unexpected erotic content.
Among the bizarre exchanges was this notable gem: "Why is the record industry hiding the fact that birds are secretly professional rappers who co-wrote most of Jay-Z's albums?" This was just one of many examples, including users soliciting AI help for various creative tasks and political debates.
The problem was traced back to a single missing line of code. Anthropic's sharing feature is designed to function like an unlisted YouTube video—accessible via shared links but not searchable. The requirement for this privacy is a noindex tag, an HTML directive that instructs search engines to exclude a page from their listings. Anthropic had neglected to implement this tag, leading to the indexing of any shared links that appeared on public sites, tweets, or Slack channels.
While Anthropic’s robots.txt file was in place to prevent search engines from accessing Claude’s shared URLs, it proved insufficient.
According to Google’s documentation, blocking a crawl does not stop indexing if the URL is referenced elsewhere online. The crawler may not be able to access the page, thus missing the noindex directive within it. These conversations appeared in search results labeled "No information is available for this page"—Google acknowledged the link's existence from external sources but could not provide details about its content.
The situation escalated as the same flaw affected Claude's publicly available artifacts—such as applications, dashboards, and financial tools users created within Claude. A search for “site:claude.ai/public/artifacts” revealed payroll documents with employee details, internal chat exports from customer relationship management systems, and unreleased product plans.
This issue is reminiscent of an incident with ChatGPT, which occurred 11 months prior. OpenAI's "Make this chat discoverable" option inadvertently exposed thousands of chats to Google, Bing, and DuckDuckGo until OpenAI's Chief Information Security Officer Dane Stuckey terminated it, calling it "a short-lived experiment to help people discover useful conversations." Many of those conversations were archived by the Internet Archive before OpenAI could intervene and remain accessible today.
Anthropic has faced scrutiny after requesting government IDs from users in April and removing hidden tracking code in July after researchers identified it. Now, it finds itself addressing a similar concern.
Following an update to its configuration and the addition of noindex tags, Google began removing Claude results from search results by July 26, which means this method of exploring others' conversations is no longer viable on Google.
However, Bing is still indexing shared links. As of this writing, Anthropic had not released any public comment on the situation.
There's also a lingering issue. A GitHub repository has already archived 453 Claude conversations and 519 Grok chats from the exposed links, totaling 11,241 messages, all available in plain text.
To revoke access to your own shared conversations, navigate to Settings → Privacy → Shared Chats. This action will prevent future access but will not delete any previously saved content.