Shared Claude chats appeared in Google and Bing search

Publicly shared Claude conversations were appearing in Google and Bing search results, WIRED reported after examining Anthropic's sharing feature. The incident did not expose random private chats or conversations that users had never shared. It did reveal a less obvious privacy risk: a link intended for selected recipients could become discoverable by anyone using a search engine.

Claude chats are private by default. When a user chooses Share, however, Anthropic creates a public snapshot with a unique URL. The snapshot contains messages and artifacts from the conversation up to that point. Anyone with the link can open it, and the owner can later remove public access. Anthropic's own help documentation describes these links as public, but many users may reasonably distinguish between ‘publicly accessible’ and ‘indexed in search.’

According to WIRED, Anthropic had blocked crawlers from the shared-chat section through its robots.txt file since at least September 2025. A sampled shared page nevertheless lacked a ‘noindex’ instruction, the more direct signal telling search engines not to list a page. Google explicitly warns site owners that robots.txt controls crawling rather than reliably keeping a URL out of search.

The practical effect was measurable. WIRED said Google had stopped showing the shared Claude pages by the time its report was published, while a Bing query still returned roughly 612 results. Anthropic said shared links are not guessable or discoverable unless users choose to distribute them, and stressed that shared content is public and can be archived. Google said publishers control which public pages can be indexed.

For Claude users, the immediate lesson is to review old links under Settings, Privacy and Shared chats, and to revoke anything that contains personal, confidential or commercially sensitive information. Shared snapshots can include more context than the final answer a user meant to show.

For AI companies and product teams, the episode highlights a broader design obligation. A secret-looking URL is not access control, and a sharing dialog should explain not only who can open a page but whether it may be indexed, copied or archived. Reliable protection needs several layers: clear warnings, ‘noindex’ metadata, sensible crawler rules and easy revocation. As workplace AI captures more sensitive context, those defaults matter as much as model performance.