Shared Claude chats turned up in search
Private Claude conversations shared via public links were found indexed on Google and Bing, exposing the gap between a robots.txt rule and real crawlers.
Over the weekend of July 25-26, people discovered that some Anthropic Claude conversations and Artifacts (the interactive documents and apps Claude can generate), shared publicly via Claude’s link-sharing feature, could be found through ordinary Google and Bing web searches, first flagged by a Reddit user. The exposed content ranged from mundane, advice on which political party to join, whether Kansas attorneys must self-report suspected ethics violations, and erotic role-play conversations, to sensitive, including resumes, financial spreadsheets, and material reported to contain crypto wallet keys. Claude lets users generate a public URL for a specific chat or Artifact to share with others, and Anthropic’s robots.txt has instructed crawlers not to index those shared URLs since at least September 2025.
The gap is a known but under-appreciated one: robots.txt is a request, not an enforcement mechanism, and both Google and Bing document that they’ll index a page anyway if it’s linked to from elsewhere on the web and the page itself lacks an explicit “noindex” HTML tag. Anthropic updated its robots.txt and took additional measures after the reports surfaced, and most flagged results had dropped out of Google by July 28. Bing was slower: at the time of Wired’s reporting, searching “site:claude.ai/share” on Bing still returned roughly 612 results.
What it means for operators
This isn’t a Claude-specific bug so much as a structural gap in how every AI chat product handles “shareable” links, and it will keep recurring across vendors until product teams treat noindex tags as mandatory, not optional, on any page meant to stay semi-private. It’s a smaller-scale version of the same lesson from Hugging Face’s agent-driven breach: the failure mode is rarely the AI model itself, it’s the ordinary infrastructure plumbing around it that nobody double-checked. If your team uses Claude’s or any competing chatbot’s share feature to hand off a conversation, a code review, a draft, anything containing information you wouldn’t want searchable, treat that shared link as public by default rather than trusting the platform’s robots.txt to keep it out of search results. The practical fix on your end is simple: don’t share anything through a public link that you wouldn’t be comfortable having indexed, and periodically search your own organization’s name alongside “claude.ai/share” or the equivalent share-domain for any other AI tool you use, to catch what a robots.txt rule alone won’t stop. That habit matters more as teams lean further into Claude’s collaboration features, including Claude Cowork’s cloud-based workspace, where more shareable artifacts naturally means more surface area for exactly this kind of exposure.