A Reddit user found sensitive data from Anthropic's Claude AI exposed through Google search, including wallet seed phrases and personal information.
A Reddit user recently uncovered a serious privacy issue with Anthropic's AI assistant, Claude, when a Google search for site:claude.ai/share revealed hundreds of private conversations. These included sensitive information such as cryptocurrency wallet seed phrases, API keys, and personal details from resumes, alongside discussions about legal obligations and explicit content.
Details of the Privacy Breach
The exposed dialogues resulted from users utilizing the "Share" function, which allows for links that can be accessed by anyone with the URL. Unfortunately, the platform did not indicate that these shared pages could be indexed by search engines, leading to significant privacy violations.
The Reddit post highlighting this issue received around 8,000 votes. The shared content was akin to "unlisted" videos on YouTube, intended to be private but susceptible to search engine indexing due to the absence of a noindex meta tag. This omission permitted Google to index shared links like any other webpage.
Although Anthropic’s robots.txt file was set up to block general chats from bots, it did not prevent indexing of shared links, as Google can still index public links regardless of crawling restrictions. This resulted in some conversations appearing in search results marked with a note that no information was available, as Google could not access the content directly.
The issue also extended to shared artifacts created within Claude, such as salary tables and unreleased product plans, which were similarly exposed through search queries.
By July 26, Anthropic had implemented noindex and nofollow tags to prevent future indexing, and Google began removing the compromised pages from its search results. Reports also indicated that links to these shared chats were emerging on Bing as well.
In a statement, Anthropic reassured users that they maintain control over their shared content and do not submit chat directories to search engines. They noted that links could only be accessed if shared by users directly.
Ongoing Accessibility Concerns
Despite the remedial actions taken, some conversations have been widely shared online, with a GitHub repository compiling an archive of 453 conversations with Claude and 519 with Grok, totaling over 11,000 text messages. While users can revoke access to their chats, this only prevents future viewing, not the visibility of previously shared content.
This incident echoes a similar mishap involving OpenAI in August 2025, when a feature unintentionally made thousands of dialogues discoverable on major search engines. The company’s Director of Information Security later termed it a "short experiment" and disabled the feature, although some chats remain accessible in the Internet Archive.
In a related note, Anthropic faced its own blunder earlier this month when a user was incorrectly billed $16.6 million on the free tier, highlighting ongoing challenges with managing sensitive user data.
