Claude's "Share" button was meant to send a link to one person. Because of a missing tag in the code, it ended up indexed by Google — and anyone could search and find conversations their owners thought were private.
On July 25, 2026, a Reddit user posted that the Google search site:claude.ai/share returned hundreds of publicly shared Claude conversations — visible to anyone, with no need for the original link. Within hours it became clear this wasn't an isolated case: there was a considerable volume of indexed chats, and someone had already scraped and uploaded more than 11,000 messages from those exposed conversations to GitHub.
The technical cause was surprisingly simple: shared conversation pages were missing a noindex tag. Anthropic's general robots.txt did exist, but that doesn't stop Google from indexing a specific page if that page doesn't carry its own "don't index" instruction — and the shared-chat pages didn't carry one. In practice, "share via a link" and "anyone can find you by searching Google" ended up being the same thing.
Google began removing the indexed results on July 26, after Anthropic added the missing noindex tags. An Anthropic spokesperson stated that "these share links aren't guessable or discoverable unless people choose to share them themselves" — a statement that's technically true but doesn't address the real problem: the missing tag is what let Google discover them without anyone actively sharing them anywhere else.
The fix stopped new indexing. It didn't erase what was already out there: direct links someone had already saved, scraped, or republished remained accessible after the fix.
This type of exposure isn't unique to Anthropic. Similar cases hit Anthropic itself in 2025, nearly 100,000 ChatGPT conversations, and xAI's Grok chatbot. The pattern repeats: a "share" feature designed for one specific use case (send a colleague a link) ends up having a much larger side effect (global public indexing) because nobody treated that page for what it actually is — public content on the internet.
If your team uses Claude, ChatGPT, or any AI assistant with a "share conversation" feature, this applies directly:
Settings → Privacy → Shared chats. Review and revoke any shared conversation that no longer needs to be public.
The incident wasn't that Claude actively "leaked" information — it was that a legitimate feature had an implementation oversight with far bigger consequences than expected. For a company integrating AI into its workflow, the lesson isn't "don't use Claude" — it's "treat every sharing feature as public from day one, because eventually it will be."
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel