Claude users' shared conversations exposed in Google search, including sensitive medical and personal data

Anthropic's Claude chatbot users discovered over the weekend that shared conversations were turning up in Google search results, potentially exposing sensitive information including patient records, company documents marked confidential, and children's names and phone numbers.

The vulnerability came to light after Reddit users found that search queries using operators like "site:claude.ai/share" were returning a vast collection of shared chats and Artifacts — interactive projects and documents built within Claude. Security researchers and journalists subsequently found health data, internal company materials, and other private content that was never intended for public viewing.

How the exposure happened#

Claude's "share chat" feature lets users generate links to conversations and projects, with the interface noting that "anyone with the link can view" the shared material. The problem arose because these links, once shared on public platforms like forums or social media, became discoverable and indexable by search engines. Unlike similar tools such as Google Docs, which have safeguards against public indexing, Claude's shared links appeared to lack comparable protections.

According to reporting from 404 Media and other outlets, exposed Artifacts contained everything from code snippets to work notes and even AI-generated erotica — the latter potentially violating Anthropic's own policy forbidding sexually explicit content creation.

Anthropic's response#

The company has attributed the issue to user behavior rather than a technical flaw. A spokesperson told TechCrunch that links only become searchable when users post them publicly themselves, and that the company does not submit share directories or sitemaps to search engines. "These shareable links are not guessable or discoverable unless people choose to share them themselves," the statement said.

However, this explanation proved somewhat contentious, since the feature's interface arguably implies the links are intended for private sharing among select recipients — similar to how Google Docs functions — rather than as tools for publishing to the broader internet.

Remediation and precedent#

By Monday afternoon, Google searches using the exposure method returned no results, suggesting the issue had been resolved. This marks at least the second time Claude conversations have been inadvertently indexed; involving roughly 600 conversations. Meanwhile, ChatGPT users have faced comparable problems — being indexed previously.