Claude shared chats leaked to Google Search for days

Thousands of Claude conversations that users had shared via public links turned up in Google search results earlier this month, exposing sexting exchanges, API keys, tax details, and proprietary code. Anthropic appears to have delisted the indexed pages over the weekend of July 26-27, but not before Reddit users spent hours cataloguing the exposed material.

The leak mirrors an almost identical incident with OpenAI's ChatGPT that Fast Company reported in July 2025. In both cases, the root cause was the same: users clicked a "share" or "publish" button without understanding that the resulting link could be crawled and indexed by search engines.

What was exposed?

Among the conversations discoverable through a basic Google query were chats containing passwords, access keys, personal questions, and what one user described as "a project meant to train future Security Engineers in Web/Cloud/AI Security." Artifacts, the mini-tools and websites Claude can generate inside a chat, were also indexed. One artifact reportedly compared two treatment arms of a medical trial and included patient names, ages, and treatment dates. Another contained what appeared to be access codes for Irish residential apartment blocks.

Fast Company chose not to publish the direct links to the most sensitive artifacts because it could not verify whether the data belonged to real individuals.

Why did this happen?

Anthropic shows disclosure language when a user clicks the share button, but the warning text is easy to miss. Rachel Tobac, CEO of SocialProof Security, called it "an obvious user experience issue." She added: "Users clearly do not understand what they are agreeing to or that they are publishing their private chats with passwords, keys, personal questions, proprietary data, tax details, et cetera to the world."

OpenAI faced the same criticism last year. Its initial defense was that the opt-in notice appeared in "smaller, lighter text in a popup." Within a week it removed the feature entirely. Google encountered a comparable problem with its Bard chatbot in 2023.

What Anthropic should do next

Tobac argues the fix is straightforward: make the publish button's consequences unmistakable, or disable the feature until users can give informed consent. "It's essential that the Claude team in charge of this function make it more clear to users that when they click that publish button, their chat will be publicly visible and indexed online, including on Google search," she said. "Or better yet, remove it until users understand and can make an informed decision that impacts their security and privacy."

Anthropic has not released a public statement explaining how long the conversations were indexed or how many were affected. The company did not respond to Fast Company's request for comment before publication.

The pattern keeps repeating

Three major AI chat providers have now exposed user conversations to search engines within three years. The underlying dynamic is consistent: users treat AI chat interfaces like private notebooks, while vendors treat sharing features like lightweight social posting. Until that mismatch is resolved, expect more leaks.

Source: Fast Company / Chris Stokel-Walker