Claude's 'Share' Feature Was Quietly Publishing Conversations to Google Search
A missing noindex tag meant shared Claude chats were publicly discoverable on Google, exposing legal advice, API keys, and crypto wallet credentials.
Over the weekend of July 26, Reddit users noticed something uncomfortable: typing site:claude.ai/share into Google returned a long list of other people’s Claude conversations. Legal discussions. Engineering work. Internal documents. At least one cryptocurrency wallet key. All sitting in search results, accessible to anyone who went looking.
This is not a story about hackers. No systems were breached. The content was technically public, in the sense that the people who created those chats had clicked a “share” button. The problem is that most of them almost certainly did not expect those conversations to show up in a Google search.
What Actually Happened
Claude’s sharing feature works the way you’d expect it to. You click “share,” a unique link is generated, and anyone with that link can view a snapshot of the conversation. The privacy implication seems obvious: if you give the link to someone, they can read it. What wasn’t obvious is what happened next.
When users posted those share links on Reddit, forums, or social media, Google’s crawlers followed the links, read the pages, and indexed them. From that point on, the conversations were searchable by anyone.
The root cause was a missing noindex meta tag on Claude’s shared chat pages. Anthropic did have a robots.txt file in place, which instructs crawlers not to visit claude.ai/share directly. But robots.txt only stops a crawler from proactively visiting a URL. Once a link to that page appears somewhere public on the web, Google can still index the page even without crawling it directly. A noindex tag embedded in the page itself is what prevents that. Those tags were absent.
The result: 11,241 shared messages were archived to GitHub before the issue gained attention. By Sunday, Google results for affected pages had largely disappeared, suggesting either deindexing or a backend fix. Anthropic later confirmed it had corrected how claude.ai/share handles crawler directives, though the company declined to specify exactly what changed, when the fix took effect, or how many conversations were indexed in total.
What Was Exposed
The exposed content was not trivial. Among the conversations researchers and journalists found indexed:
- Legal advice and personal legal discussions
- Corporate documents and internal employee reviews
- Resumes and professional data
- Proprietary code and engineering work
- API keys
- At least one cryptocurrency wallet private key
That last item deserves separate attention. A leaked private key is not a recoverable situation. There is no support ticket, no fraud department, no way to reverse a transfer. If someone found that key and moved funds before the chat was deindexed, those assets are gone permanently.
Claude Artifacts, the interactive mini-apps and documents users can build inside Claude, were also affected. Shared artifacts were indexed alongside standard conversations.
Anthropic’s Response
Anthropic’s public position, delivered via spokesperson Amie Rotherham to TechCrunch, was that share links “are not guessable or discoverable unless people choose to share them themselves.” The company pointed out that links only appear in search results when they’ve been posted somewhere public.
That framing is technically accurate and practically misleading. Users who shared a link in a forum thread to get help with a problem were not expecting Google to index the contents of that conversation. The reasonable expectation when someone clicks “share with a link” is that the audience is whoever they send the link to, not the entire internet.
Anthropic has not published a technical description of what was changed, confirmed a patch date, or acknowledged how many users were affected.
It’s also worth noting this is not the first time this has happened. A similar incident in early 2025 affected roughly 600 conversations under comparable circumstances. A comparable situation also arose with ChatGPT share links. The same missing control, repeated across multiple platforms, points to a broader pattern in how AI products handle the privacy implications of convenience sharing features.
What You Should Do Right Now
If you have ever used Claude’s share feature, it is worth spending five minutes checking your exposure.
To review your shared chats: Go to Settings, then Privacy, then Shared Chats. This shows every conversation you have set to public.
From there, you can unshare conversations individually. Doing so immediately revokes access via the shared link. The content is no longer viewable by anyone who had the URL.
If your shared conversations contained anything sensitive, including work documents, personal data, financial information, or anything you would not want indexed on Google, unsharing now prevents further access. It does not, however, undo whatever happened before Sunday. Anyone who bookmarked or saved a URL before Anthropic’s fix still has a working link unless access is revoked server-side.
If you shared a cryptocurrency wallet key or seed phrase in a Claude conversation at any point, assume that information is compromised and move assets to a new wallet immediately.
The Broader Point
This incident is a good reminder that “share with a link” and “searchable by anyone” are closer together than most people assume. Share links are convenient precisely because they are easy to distribute, but that same property makes them easy to leak. Once a link touches the public web, you have limited control over where it ends up.
For professionals using Claude for work, the risk extends beyond embarrassment. Shared conversations containing proprietary code, client information, or business data could create intellectual property or compliance problems. Enterprise and Team plan users are somewhat better protected, as sharing in those tiers is restricted to members of the same organisation. But free and Pro users on individual plans had no such guardrail.
Until AI platforms consistently apply noindex tags, authentication requirements, or expiring links to shared content, the safest approach is to treat any shared AI conversation the way you would treat a document posted to a public website. If you would not be comfortable with it turning up in a Google search, do not share it with an open link.