Anthropic's Claude Share Links Indexed by Google and Bing

A share button that quietly turned into a publish button. That is the short version of what happened this week to Anthropic's Claude, and it is the second time in about a year the company has hit the same failure mode.

According to Wired, users on Reddit and X found they could pull up shared Claude conversations and Artifacts by searching Google and Bing. The exposed material was not thin. Reporters and researchers surfaced what appeared to be a real patient's medical report, clinical trial results that included patient names, documents containing the names and phone numbers of primary school aged children, employee reviews with personal information, and company documents marked for internal use only. When a Claude user clicks Share, the app generates a public URL at claude.ai/share followed by a unique identifier. Those pages were reachable by anyone with the link, and by any crawler that could find one.

The technical mistake will feel familiar to anyone who has run a website. As Search Engine Journal walked through, a robots.txt entry only asks a crawler not to fetch a page; it does not tell a search engine to keep the URL out of its index if the link is discovered elsewhere on the open web. To do that you need a noindex header or meta tag on the page itself, and Anthropic's shared chats did not have one. Google began pulling Claude results by July 26 after Anthropic added noindex tags, per TechCrunch. Anthropic's own line is that "When someone shares a conversation, they are making that content publicly accessible, and like other public web content, it may be archived by third-party services," which is technically true and also a pretty telling reframe of what most users probably thought Share meant.

The honest caveats are worth stating. No one has published a full count of how many conversations were indexed this time. Last year Forbes reported a similar issue at Anthropic, and Google's estimate then was just under 600 conversations indexed before the pages disappeared. The reporting also does not tell you whether third-party archives such as the Internet Archive or Bing's cache still hold copies, or whether any of the exposed material has been scraped into training data. Those secondary questions are worth watching more than the immediate cleanup, because a page that was public for a week can outlive a fix by years.

For everyone else building consumer AI products, this is a cheap lesson bought with someone else's users. A Share button on sensitive content needs noindex on the shared page by default, an unambiguous warning that share means the open web, and probably an expiry option too. The companies that get this right first turn a recurring embarrassment for the category into a quiet trust advantage.