Anthropic’s Claude accidentally put other people’s shared chats on Google — and some of them included sensitive crypto data.
What happened
- Every Claude shared link comes labeled “Anyone with the link.” But Reddit user sleuthed that Google was indexing those links. A site-specific search (site:claude.ai/share) run on July 25 returned hundreds of people’s conversations and the Reddit thread scored ~4,000 upvotes.
- The indexed pages included everything from mundane chats to deeply sensitive items: a user pasting a crypto wallet seed phrase (the master key to a wallet), legal questions about ethics violations, internal spreadsheets, product roadmaps, and bizarre or explicit queries. One memorable entry: “Why is the record industry hiding the fact that birds are secretly professional rappers…?”
- The root cause was one missing line of HTML: Claude’s share pages lacked a noindex tag. Anthropic’s share feature was meant to act like an unlisted video — link-only access — but without noindex, search engines could index any URL they learned about elsewhere on the web.
- Anthropic did have a robots.txt rule blocking Claude share URLs. That wasn’t enough: Google’s crawler can still index a URL if it finds the link on other sites, even when it can’t fetch the page to read an internal noindex tag. The search results showed “No information is available for this page” while still linking to the exposed content.
- The exposure extended beyond chats to published “artifacts” — payroll spreadsheets, CRM exports, unreleased product roadmaps — via site:claude.ai/public/artifacts queries.
- Google began removing Claude results by July 26 after Anthropic added noindex tags and updated configuration. Bing, however, is still surfacing some shared links.
- A public GitHub repo named Shared-Claude-Chats has already archived 453 Claude conversations and 519 Grok chats — 11,241 messages in plain text — meaning copies of exposed chats persist outside Claude’s control.
- Anthropic had no public statement at the time of publication. This incident follows other recent Anthropic hiccups: asking users for government IDs in April and quietly removing hidden tracking code in July after it was discovered.
Why crypto users should care
- Wallet seed phrases or private keys pasted into AI chats are a catastrophic single point of failure. Anyone who finds a seed phrase can fully control the associated funds. If you ever pasted a seed phrase, private key, or other credentials into a Claude share (or any cloud chat), treat the key as compromised.
What to do now
- Revoke future access: In Claude, go to Settings → Privacy → Shared Chats and revoke any shares. That stops future access but does not delete copies already archived elsewhere.
- If you exposed a wallet seed/private key: move funds immediately to a new wallet with a fresh seed (preferably using a hardware wallet). Assume any pasted credentials are compromised.
- For teams: audit shared artifacts (spreadsheets, exported chats, dashboards) and rotate any exposed credentials, API keys, or passwords. Remove or re-host sensitive docs that were shared via Claude.
- General best practices: never paste seed phrases or private keys into cloud services; use hardware wallets and offline key storage; treat any third-party share link as potentially public until proven otherwise.
Context and precedent
- This is not new terrain: OpenAI had a similar problem about 11 months earlier when a “Make this chat discoverable” setting allowed thousands of chats to be indexed and some ended up archived in the Internet Archive before being cleaned up.
- Unlike a simple UI switch, the Claude issue was a small but crucial missing HTML instruction — showing how tiny configuration errors can expose sensitive material at scale.
Bottom line
Anthropic fixed the immediate indexing issue for Google by adding noindex tags, but copies of exposed chats already exist in third-party archives and some search engines still surface links. If you’ve ever shared sensitive crypto credentials in Claude (or similar services), act now: revoke shares, rotate keys, and assume compromise.
Read more AI-generated news on: undefined/news
