Shared chats and Artifacts for which users had created public links in Anthropic's AI service "Claude" were found to be collectively discoverable through Google search results. A Reddit user discovered this on July 25, and on the 27th, Futurism confirmed pages containing medical information and internal corporate documents. The search results subsequently disappeared, but known shared URLs remained accessible. This issue is not a case of intrusion into private chats. Rather, it lies in the design of the sharing feature—specifically, how discoverable a URL that a user has voluntarily made public should remain.
Which Chats Appeared in Search Results
The issue originated from a Reddit post noting that searching Google for site:claude.ai/share returned numerous Claude shared pages. The exact number was not confirmed. TechCrunch reported that on the afternoon of July 27, when it tried the same search, no results were returned. Fortune also confirmed the disappearance from search results, but noted that people who already knew the URLs could still access the shared pages.
What was affected were chats for which users had created public URLs through Claude's sharing feature. According to Anthropic's official help documentation, regular chats are private by default. When sharing is executed, a snapshot is created containing the messages and Artifacts up to that point, and anyone who knows the URL can view it without authentication. Conversation that continues after sharing is not automatically added. Once sharing is revoked and then re-enabled, the snapshot is updated to include the new messages.
Therefore, not all of a user's Claude chats were passed to search engines. Nor has it been confirmed that public links allow access to an entire account or to other conversations. However, once a URL that a user intended to send to a specific recipient becomes searchable, the range of viewers changes entirely. Neither the total number of pages indexed this time nor where search engines discovered each URL has been disclosed.
From "Anyone with the Link" to Searchable Pages
Two distinct states of publicness overlap in shared pages. One is "accessibility"—the ability to open a URL without authentication. The other is "discoverability"—whether a search engine finds the URL and lists it in search results. Long, random URLs make brute-force guessing difficult, but if a recipient posts the link on social media or a forum, or if it ends up in a public repository or web log, search crawlers can trace it through external links.
Anthropic told TechCrunch and others that it does not provide a directory or sitemap of shared chats to Google, and that shared URLs cannot be guessed or discovered unless users share them externally themselves. Google responded that the affected pages were indexed by multiple search engines, and that site operators have means to control crawling and search registration. Both companies' explanations are compatible with each other, since search engines can find pages starting from publicly shared links even without receiving a list of URLs.
Google's official documentation states that pages one does not want appearing in search results should use authentication or noindex. Even if crawling is disallowed via robots.txt, Google may still list a page in search results with limited information if it discovers the URL through an external page. For sensitive information, access controls such as passwords are necessary; noindex does not prevent viewing by anyone who already knows the URL.
As of July 28, Claude's robots.txt disallows crawling of /share/*. HTTP responses for shared pages also return X-Robots-Tag: none. According to Google's specifications, none is equivalent to noindex, nofollow. However, it remains unclear when this header was added, or which method was used to remove the pages from search results. Google also explains that for a crawler to read a page's noindex directive, it must first be able to access the page—so how this current dual configuration is handled by each search engine still needs ongoing verification.
The Same Explanation as 2025, When Nearly 600 Cases Were Counted
This is not the first time Claude's shared chats have appeared in search results. Forbes reported on September 8, 2025, that Google was estimated to have indexed nearly 600 conversations. At that time as well, Anthropic explained that the chats became visible on Google and Bing because users had posted the links online or on social media, and stated that it had not provided a directory or sitemap of shared pages.
However, one of the users identified by Forbes in 2025 said they had not posted online the work-related chat that appeared in search results. This time as well, Anthropic has repeated the explanation that shared URLs cannot be discovered unless users share them themselves. This is not evidence that Anthropic directly provided the URLs to search engines. Conversely, it also does not prove that users' public posts alone can account for how every URL was discovered.
While Google's previous estimate was under 600 cases, this time the total number is unknown, making it impossible to compare the scale of the two incidents. Nevertheless, essentially the same corporate explanation for a similar type of exposure has resurfaced between September 2025 and July 2026. The stopgap measure of removing pages from search results and the design of communicating the meaning of the sharing action to users must be examined separately.
Differing Warnings for Artifacts and Chats
Futurism opened shared pages from Google search results and confirmed detailed medical reports apparently belonging to real patients, clinical trial results including patient names, and documents listing the names and phone numbers of elementary-school-age children. The outlet also reported finding corporate documents marked internal-only and evaluation materials containing employees' personal information. To protect the privacy of those involved, the outlet did not publish the URLs in question.
There were also differences in the warnings given at the time of publication between chats and Artifacts. According to Futurism, the chat sharing screen creates a "public link" and states that "anyone with the link can view it," but did not clearly indicate the possibility of registration with search engines. The Artifact publishing screen, on the other hand, warned that anyone on the internet could access it and that it might appear in search results. Even within the same Claude product, the scope of publicness users assume can differ depending on the feature.
Reading the official help documentation reveals even finer distinctions in the boundaries of shared data. The actual files attached to a chat are not included in the snapshot, and raw data retrieved through MCP integration is also hidden. What viewers see is the conversation and Claude's responses. However, if the content of an attachment or MCP data is pasted into the conversation body, or if Claude writes it out in its final response, that text can remain in the shared snapshot. Even when the attached file itself is excluded, information that has been expanded into the conversation remains.
There are also differences by plan. Public sharing of chats can be managed by users on the Free, Pro, and Max plans, while Team and Enterprise users can only share chats with members of the same organization. To determine the scope of publicness, one must check under which subscription and account the sharing was performed.
Stop the Shared URL Before Relying on Search Result Removal
Free, Pro, and Max users can view a list of their shared chats' titles, share dates, and URLs from Settings > Privacy > Shared chats in Claude. Unwanted shares can be disabled via Unshare. The absence of results in Google search is not proof that a shared page has been made private. Since known URLs remained accessible even after being excluded from search, checking Claude's own shared-chat list should be the first step.
If a URL containing sensitive information was made public, third-party archiving or reposting may persist even after revocation. Anthropic itself explains that published shared pages can be archived by third-party services. For documents intended to be shared only with specific recipients, it is better to move them to storage with authentication and permission management, rather than using Claude's public links as a distribution channel.
On the service side, there is a need for mechanisms that clearly indicate discoverability at the moment of sharing, and that separately control public links versus search registration. If exclusion from search is intended, noindex must be maintained in a form readable by each search engine, and the status of past URL removals must also be tracked. This issue can only be considered resolved not when search results temporarily go empty, but when users have been informed about the warnings on the sharing screen, the handling of existing URLs, and the methods for exclusion across multiple search engines.