Hundreds of conversations people had with Anthropic's Claude chatbot were found available to anyone searching online. Links to these chats showed up through simple search engine queries.
The exposure happened when users chose to "share" a conversation link. Those links were then picked up and saved by search engines like Google.
Once saved, the chats appeared to anyone who searched for them. This included people who had no connection to the original conversation.
Anthropic removed the chats from search results over the weekend. Many of the conversations had already been saved and shared widely elsewhere online.
How the Chats Were Found
Users on Reddit first discovered the exposed chats. They found more than 200 conversations spread across at least 25 pages of search results.
Some of the chats were recent, taking place just weeks before they were discovered. Others were from as far back as last year.
The conversations covered a wide range of topics. Some included personal details such as names, contact information and work history taken from CVs.
Other chats contained research tied to people's jobs, including healthcare related material. Some even included transcripts of private conversations.
One chat from April showed a user asking Claude to draft an unpublished blog post. The post included details of a corporate cloud security project.
In another exchange, a user asked Claude how to turn into a nine tailed fox. Claude responded by generating an image and telling the user they now had "fully functional fox powers."
A separate chat showed a user asking Claude directly whether it wanted to help the user or help Anthropic more. Claude replied that it experienced something like wanting to help the user.
Anthropic's Response
A spokeswoman for Anthropic said users control if and when their conversations get shared. She said links to conversations are not guessable unless someone chooses to share them.
She added that once a person shares a conversation, that content becomes public. Like other public web pages, it may then get archived by outside services.
The Claude share feature tells users that anyone with the link can view the conversation. It does not clearly state that shared links could end up listed in search results.
A Google spokesman told the BBC that the company does not control which pages are made public on the web. He said that action comes from the website owners themselves.
Google gives site owners tools to decide whether pages can be crawled or indexed. The spokesman said Google always respects those settings.
This is not the first time an AI chatbot has faced this kind of exposure. OpenAI's ChatGPT had a similar issue last year, and the company later changed how shared chat links could be accessed.
Grok, the chatbot built into Elon Musk's platform X, also had hundreds of thousands of chat logs appear in search results last year.
Since the exposure was removed from search engines, Anthropic appears to have used available tools to block the chat links quickly. Other search engines, including Bing, Brave and DuckDuckGo, where the Claude chats also appeared, were approached for comment.
