Headline: Claude share links were quietly indexed by Google — and exposed seed phrases, payrolls and more
A simple Reddit search on July 25 pulled back a troubling discovery: hundreds of Claude.ai “shared” conversations were showing up in Google results. Every share link Claude generates is labeled “Anyone with the link,” and because Anthropic didn’t include a single HTML noindex tag, those links—when posted publicly on forums, Twitter or Slack—became discoverable by search engines. A Reddit thread about the find racked up 4,000 upvotes and offered a lurid snapshot of what people will tell an AI: a lawyer worrying about an ethics violation, someone pasting a crypto wallet seed phrase (the master key to funds), payroll spreadsheets with employee names, internal CRM exports, unreleased product roadmaps, weird creative roleplay, and everything in between.
How this happened
- Anthropic’s share feature functioned like an unlisted YouTube link: intended to be accessible only by recipients, but not indexed. To actually prevent indexing the page must include a noindex meta tag. Anthropic hadn’t added it.
- Anthropic’s robots.txt was already set to block the share URLs, but that alone isn’t enough. Google’s crawler can index a URL it finds elsewhere on the web without crawling the page. When that happened, Google listed the links with the label “No information is available for this page” — it knew the URL existed from third-party posts but couldn’t see inside the page to obey any noindex instruction.
- The same oversight exposed larger “published artifacts” (site:claude.ai/public/artifacts) — payrolls, internal chat exports, roadmaps — that people had built inside Claude.
Context and precedent
- This isn’t the first time AI chat shares accidentally hit search results. About 11 months ago, OpenAI’s “Make this chat discoverable” experiment sent thousands of ChatGPT conversations to Google, Bing and DuckDuckGo until OpenAI pulled the feature; many of those chats were archived by the Internet Archive before they could be removed.
- Anthropic has had other recent privacy stumbles too: in April the company asked some users for government IDs and in July quietly removed hidden tracking code from Claude Code after researchers flagged it.
What changed and what remains exposed
- Anthropic updated its configuration and added noindex tags, and Google began pulling those Claude results by July 26. Anthropic had not issued a public statement by the time of publication.
- However, Bing is still returning some shared Claude links, and a GitHub repo called Shared-Claude-Chats has already archived 453 Claude conversations and 519 Grok chats — 11,241 messages in total — pulled from exposed links. Public archiving means revoking a link won’t unsave copies that others have already archived.
Why crypto users should care
- The most alarming discoveries were directly relevant to crypto security: plain-text wallet seed phrases, account credentials and other highly sensitive material appeared in exposed chats. Anyone who pasted a seed phrase, private key, or account password into a shared Claude chat and then posted that link publicly (or had it reposted) may have had those secrets exposed and/or archived.
Quick steps to protect yourself
- Revoke shared links: In Claude go to Settings → Privacy → Shared Chats and turn off or revoke shares. That prevents future access but won’t erase copies already archived elsewhere.
- Rotate keys and credentials immediately if you pasted them into a shared chat that might be public.
- Stop pasting private keys, seed phrases, or passwords into AI chats or any chat you might share.
- Search for your own shared links (site:claude.ai/share and site:claude.ai/public/artifacts) and check other search engines and public archiving sites.
Bottom line
A single missing line of code — the noindex tag — turned “anyone with the link” from a reasonable privacy model into a mass leak when links were redistributed. The incident is a reminder that default sharing settings and modest implementation oversights can create outsized risk, especially for crypto users who may inadvertently expose seed phrases or account data in ostensibly private AI conversations.
Read more AI-generated news on: undefined/news
