A developer pastes an API key into Claude, shares the conversation with a colleague, and assumes “anyone with the link” means exactly that — someone who already has the link. Months later, that conversation is sitting on Google for strangers to find. That’s reportedly what happened to roughly 600 Claude conversations indexed by Google and Bing, discovered on July 26, 2026. This wasn’t a hack. Nobody breached Anthropic’s servers. The share feature worked exactly as designed — which turned out to be the problem.
What Was Actually Exposed
The indexed conversations reportedly contained far more than casual prompts.
Reported content found across the searchable chats included:
- Crypto wallet keys, API credentials, and login tokens
- Legal strategy discussions and internal company documents
- Resumes containing names, addresses, and phone numbers
- An AI-powered therapy app prototype and a medical billing dashboard
- Meeting notes with confidential business details
Claude’s share feature creates a public webpage snapshot of your conversation. You choose “public” or “anyone with the link,” and that snapshot captures everything typed before the share point. Anthropic reportedly said it does not provide sitemaps or directories to search engines and actively blocks crawling. But if a shared link lands somewhere crawlable — a forum post, a public Slack, a tweet — search engines do exactly what they’re built to do.
Think of it as leaving your unlocked phone on a café table in Times Square, then blaming tourists for looking at the screen.
This pattern has precedent. ChatGPT reportedly had a similar episode last year, with nearly 100,000 conversations becoming searchable on Google. The Google query technique that surfaced Claude chats was reportedly mitigated, though the Artifact-related search query was still working at time of reporting.
Revoking a Share Doesn’t Erase the Trail
Removing a public link only controls what Claude serves going forward — not what’s already been copied or cached elsewhere.
If a third party archived or saved your shared page before you revoked it, that content still exists somewhere outside your control. You can review and unshare conversations inside Claude’s Privacy settings. That step is necessary — but it’s not sufficient.
Every AI chatbot with a share feature carries this failure mode baked in. The feature was marketed as shareable. It was. Just not in the way you — or apparently Anthropic — intended. Your shared chat isn’t a DM. It’s a webpage. Treat it like one. Just as a surveillance app can expose data in ways users never anticipated, any tool designed to share can become a liability when its defaults aren’t fully understood.
