On July 25, 2025, a serious security vulnerability was discovered on Reddit in the AI assistant Claude by Anthropic. A user under the nickname fuadmefleh found that the query site:claude.ai/share in the Google search engine returned hundreds of other users' conversations with the neural network. Among them were seed phrases from crypto wallets, API keys, resumes with personal data, and social security numbers.

The problem lies in the "Share" feature in Claude. It creates public links to chats that, by design, should remain hidden from search engines, similar to "unlisted" videos on YouTube. However, Claude lacked the noindex meta tag, which prevents page indexing. As a result, any published conversation automatically ended up in Google's index.

Why did this happen?

Anthropic's robots.txt file was configured correctly and blocked crawlers. But, as experts explain, this does not prevent indexing if a link to the page already appears in public sources. A crawler that is blocked from access cannot see internal meta tags, such as noindex. Thus, some conversations appeared in Google search results with the note "no information about the page" — the search engine recorded the URL but did not read the content.

The same fate befell Claude artifacts — applications, dashboards, and documents that users create within the platform. The query site:claude.ai/public/artifacts revealed salary tables with names, exports from corporate CRMs, and unreleased product plans.

Anthropic's response and consequences

By July 26, Anthropic had added the noindex and nofollow meta tags. Google began removing pages from search results. However, according to Decrypt, links to shared chats were also shown by Bing. Anthropic representatives stated that users themselves control the publication of conversations, and the company does not provide search engines with directories or sitemaps of chats. They emphasized that links cannot be found unless a person has shared them themselves.

Despite this, some conversations have already been copied. A repository called Shared-Claude-Chats appeared on GitHub, containing an archive of 453 conversations with Claude and 519 with Grok, totaling 11,241 text messages. Access can be revoked in Claude's privacy settings, but this only prevents further viewing, not the deletion of already copied data.

Similar incidents

This is not the first case of data leakage through AI assistants. In August 2025, OpenAI accidentally indexed thousands of ChatGPT conversations through the "Make this chat searchable" feature. The company's Chief Information Security Officer, Dane Stuckey, disabled the option, calling it a "short experiment." Some of those chats are still available in the Internet Archive. Earlier, Google indexed user conversations with DeepSeek, including work documents and personal correspondence in Russian.

Expert opinion: This incident highlights a fundamental security problem in modern AI platforms. Developers often neglect basic privacy principles, relying on users for data control. In the world of cryptocurrencies, where seed phrases are keys to assets, such vulnerabilities can lead to irreversible financial losses. Anthropic and other companies should implement automatic checks for sensitive data before publication, rather than relying on post-factum fixes.