Last week, if you knew what to search for, you could read a stranger’s medical report on Google. Next to it, clinical-trial results with real patients named. Elsewhere, a list of primary-school children with their phone numbers, internal company files, and a set of employee performance reviews with people’s names attached.
This was not a hack. Every one of those documents was put there by the person it belonged to, or by someone they trusted, who typed it into Claude and pressed a button marked “share“.
Redditors were first to raise the alarm. Google started scrubbing the pages around 26 July, roughly 600 of them. Bing was reportedly slower.
One Button, Two Ideas of “Share”
Claude’s share button does exactly what it says: it spins your conversation into a public web page and warns you, in plain text, that anyone with the link can view it.
However, as we are finding out, this is not as simple as it may seem.
Users are reading this and then picturing sharing the link to a colleague. It appears Google has a different idea, as it reads that same page and sees an ordinary website begging to be indexed.
This Keeps Happening
It’s easy to dismiss this as ‘another silly little Claude quirk’. After all, is it that big of an issue to see a list of primary-school children with their phone numbers? Or someone’s medical report? Or internal company files? Right? But the even bigger issue here is that the button that exposed it is not unique to Claude. This is the fourth time a chatbot has done this in a year.
In July 2025, OpenAI had to yank a ChatGPT feature that let users make shared chats discoverable, after thousands showed up in search. Weeks later it was Grok’s turn, with no warning at all and hundreds of thousands of transcripts exposed, some of them genuinely deranged, including a Grok-written plan to assassinate Elon Musk. Last September, Forbes found around 600 Claude chats sitting in Google.
Four leaks, three companies, twelve months. Every major lab has shipped the same tempting one-click share, and every time, a chunk of users have used it to publish things they definitely don’t want on the internet.
It’s not looking that promising now, is it?
Anthropic’s Shrug
Anthropic’s answer is technically airtight, but besides the point.
It does not hand Google a directory of everyone’s shared chats, and the links are not guessable, so nothing surfaces unless a user chooses to share it and then, usually, posts it somewhere a search crawler can find. All true. However, none of it actually helps the person whose oncologist’s report is now cached across the internet.
And if that spreadsheet of employee reviews happened to belong to your company, there is another problem. The leak is not Anthropic’s fault. No, it’s yours.
See Also:
Biggest AI Surveillance Scandals Threatening Europe’s Privacy in 2026
Is Claude Safe for Production? AI Agent Deletes Database in 9 Seconds
Company Blew $500M on Claude Because Nobody Set a Spending Limit
