The AI User Nightmare: Private Claude Conversations Leaked to the Internet
Hundreds of Anthropic Claude chat logs, containing medical records and internal documents, appeared in Google search results after being shared by users.
A significant privacy breach has emerged involving hundreds of private conversations with Anthropic's popular AI chatbot, Claude.
Recent findings reveal that these chat logs became accessible to the general public via standard Google searches, exposing sensitive personal and professional information.
The vulnerability arose from a feature allowing users to "share" their chats.
While intended for sharing with colleagues or friends, these shared links were indexed by search engines.
Users on Reddit discovered that by employing specific search operators—specifically searching 'site:claude.ai/share'—they could uncover hundreds of conversations dating back several weeks.
The exposed data included highly sensitive content, such as medical records, clinical trial results identifying patients, internal corporate documents marked "for internal use only," and personal contact details, including names and phone numbers of children.
Beyond medical data, the leak included professional information such as resumes containing full work histories and contact info, as well as 'Artifacts'—the interactive mini-applications built within the Claude interface.
Some logs even contained employee performance reviews containing private individual details.
Anthropic has since moved to prevent further indexing of these links.
Amy Rotherham, a spokesperson for Anthropic, stated that the company removed the searchability of chat conversations immediately upon receiving reports of the issue.
However, she noted that much of the content had already been widely shared across the web.
Rotherham emphasized that Claude conversation links are not guessable or discoverable unless a user explicitly chooses to share them, adding that once shared, such content becomes public and can be cached by external services.
While Claude's sharing interface does warn users that "anyone with the link" can view the content, critics argue the phrasing implies sharing within small groups rather than the entire internet.
Notably, the interface does not explicitly warn that these links may be indexed by Google.
Google spokesperson Ned Adriance clarified that, like all search engines, Google does not control the content of web pages but respects the decisions of site owners regarding crawling and indexing.
This incident is part of a recurring pattern in the AI industry.
Last year, similar issues were reported involving nearly one hundred thousand Claude conversations being indexed, as well as public access to chat histories from ChatGPT and Grok.
As users increasingly rely on AI as a space to "think out loud" about sensitive work, legal, and health matters, the industry faces a growing challenge in securing these highly personal data streams against unintended public exposure.