On July 25, a Reddit user discovered that searching for site:claude.ai/share in Google yielded hundreds of private conversations with Anthropic's AI assistant, Claude. Among these were sensitive items like cryptocurrency wallet seed phrases, API keys, resumes containing names and phone numbers, and numbers resembling U.S. Social Security numbers.
Source: Reddit.How the Links Ended Up in Google Search
The dialogues that appeared in search results were exclusively those shared by users through the "Share" function, which creates a link that can be accessed by anyone with the URL. However, the interface did not warn users that these pages could be indexed by search engines.
The Reddit thread garnered approximately 8,000 votes. In addition to the seed phrases and access keys, the indexed chats included a lawyer's question about whether he was obligated to report ethical violations, discussions about policies, and explicit content.
The "Share" feature in the chat bot functions similarly to "unlisted" videos on YouTube: while users can share the link, it should not be discoverable through searches. This is typically managed by a noindex meta tag, which prevents search engines from indexing the page. However, Claude lacked this tag, allowing any shared link to be indexed like a regular webpage.
Moreover, Anthropic's robots.txt file was correctly configured to block general chats from bots. However, this was insufficient. According to Google’s logic, a prohibition on crawling does not stop indexing if the link is available in public sources. A bot that is denied access cannot see the internal noindex tag. Consequently, some dialogues appeared in search results with a note indicating that no information about the page was available — the search engine detected the address from external publications but could not read the content.
This issue also affected shared artifacts — applications, dashboards, and documents users compiled within Claude. A search for site:claude.ai/public/artifacts revealed salary tables with names, corporate CRM exports, and unreleased product plans.
By July 26, Anthropic implemented noindex and nofollow tags. Google began removing the pages from search results. Observations by Decrypt noted that links to shared chats were also appearing on Bing.
In a statement to VentureBeat, representatives from the AI developer indicated that users have control over their chat publications, and Anthropic does not submit directories or sitemaps of chats to search engines. They emphasized that links cannot be discovered unless shared by users themselves.
What Remains Publicly Accessible
Some of these conversations have already been widely disseminated. A repository called Shared-Claude-Chats appeared on GitHub, containing an archive of 453 conversations with Claude and 519 dialogues with Grok, totaling 11,241 text messages. Users can revoke access in Claude's settings under the privacy section, but this only prevents further viewing.
A similar incident occurred with OpenAI in August 2025, when the "Make this chat discoverable" feature inadvertently sent thousands of dialogues to Google, Bing, and DuckDuckGo. The company’s Director of Information Security, Dane Stakey, disabled the option, calling it a "short experiment." Some of those chats are still accessible via the Internet Archive.
Previously, Google indexed dialogues from users of the DeepSeek AI assistant, which included working documents, financial analyses, and personal communications, including those in Russian.
It is worth noting that in July of this year, Anthropic mistakenly billed a user on the free tier for $16.6 million.