Claude AI data leak: Private chats were freely accessible on Google

Due to an insufficiently protected sharing function, private chat logs from the Anthropic AI Claude leaked onto the public network. Web crawlers captured the URLs and listed medical reports and personal information in search engines.

Sensitive AI conversations online

Many users of the Claude language model have unknowingly shared confidential information with the general public. This past weekend, private chat logs and programmed applications began appearing in major search engine results. The reason was the sharing function of the AI ​​platform, which is actually intended for sharing with colleagues or friends. According to the warning from the provider Anthropic, anyone who generates a link to a conversation makes the document accessible to anyone with the link. Many users overlooked the risk that web crawlers could capture these unprotected URLs. The range of leaked content ranges from harmless notes to critical data.

The data was loud 404 Media tracked down using specific search queries. In addition to pure text chats, so-called artifacts are also affected. These are small programs and tools generated by artificial intelligence that users can run directly in the browser. Among the documents found were medical patient reports, internal company evaluations and private crypto keys. Personal data such as names, telephone numbers and addresses of children were also freely available online. Even erotic role-playing conversations that actually violated the terms of use appeared in the search results.

Lack of technical protection

An advantage of the sharing function is the easy collaboration on projects without additional software. The disadvantage is the inadequate technical security. Anthropic relies on a simple text file to stop search engines from indexing. More robust protection through specific HTML tags is completely missing from the generated websites. When shared links are posted on public forums or social networks, many search engines ignore the rudimentary blacklists and include the pages in their index. This means that conversations that were only intended for a small group of people suddenly become discoverable and permanently archived for the entire world.

The question of guilt remains unclear

The developer sees the responsibility as the users. Anyone who actively shares content makes it publicly accessible. Search engine operators, on the other hand, emphasize that they only display the visible network. They point out the obligation of website operators to protect their pages from being read with the correct technical markings. Most of the affected search results now appear to have disappeared. Users can review and customize their sharing to maintain privacy.

To do this, they need to open the general settings, go to the privacy section and manage or delete shared conversations. In order to best protect sensitive data, confidential texts should be copied manually and sent via secure channels. The problem is not new, as similar incidents occurred on various services last year.