A simple search with the parameter “site:claude.ai/share” returned thousands of chats from Claude users over the weekend. These also contained confidential data and information. Google no longer displays entries found with this search.
A user on Reddit discovered the Claude problem. In the course of the discussion, participants also found chats with sensitive information. For example, crypto keys for wallets were found there, or a lawyer's request as to whether he should report himself for misconduct. It is currently unclear to what extent private chats became public as a result. Although Google no longer shows results with the search term, Reddit users report that initially Bing or the Brave browser search still found the results – however, we could not verify this in a brief check at the time of reporting.
A known problem
Apparently, the cause of the problem is the same one that OpenAI experienced with ChatGPT in August of last year: there is a function for sharing chats that creates a public link. In addition to link creation, there was an option “Make this chat discoverable”, meaning “mache diesen Chat auffindbar”. The option was apparently often misunderstood at the time. Anthropic offers privacy settings under “Settings” – “Privacy” – “Shared Chats” or “Shared Artifacts”; the shared resources can be found there. In response to a request from heise online, Anthropic explained that the sharing links are only discoverable if users share them publicly themselves.
