Google Leak: Seed phrases and API keys found in published Claude chats
On July 25, a Reddit user discovered that the search query site:claude.ai/share returned hundreds of other users' conversations with Anthropic's AI assistant, Claude. Among the data indexed by Google were a cryptocurrency wallet seed phrase, API keys, resumes containing personal information, and numbers resembling U.S. Social Security numbers.
How the links ended up in search results
Only those conversations that users had published via the "Share" feature appeared in the search results. This option creates a link accessible to anyone with the address, but the interface did not warn that the page could be indexed by search engines. The mechanics are similar to "unlisted" YouTube videos: the link can be shared, but the video should not be found through search. This is controlled by the noindex meta tag, which was absent in Claude. As a result, any published address automatically became a regular, indexable page.
Anthropic's robots.txt file was configured correctly and blocked general chats from bots. However, this proved insufficient: according to Google's logic, a crawl ban does not prevent indexing if the link appears in public sources. A bot that is blocked from access also cannot see the internal noindex tag. Therefore, some conversations appeared in search results with a note indicating a lack of page information — the search engine recorded the address from third-party publications but did not read the content.
The same issue affected published artifacts — applications, dashboards, and documents that users compile within Claude. Searching for site:claude.ai/public/artifacts revealed salary tables with names, exports from corporate CRMs, and unreleased product plans. By July 26, Anthropic had added noindex and nofollow tags, and Google began removing pages from search results. Links to shared chats were also displayed by Bing.
What remains publicly accessible
Some of the conversations have already been distributed in copies. A repository called Shared-Claude-Chats appeared on GitHub, containing an archive of 453 conversations with Claude and 519 dialogues with Grok — totaling 11,241 text messages. Access can be revoked in Claude's settings under the privacy section, but this will only block further viewing.
A similar incident occurred with OpenAI in August 2025, when the "Make this chat searchable" feature sent thousands of conversations to Google, Bing, and DuckDuckGo. The Chief Information Security Officer disabled the option, calling it a "short experiment." Some of those chats are still available on the Internet Archive. Previously, Google also indexed user conversations with DeepSeek, including work documents and personal correspondence.
Expert opinion: This incident is yet another reminder that even major AI platforms do not always ensure an adequate level of privacy by default. For the crypto community, it is particularly alarming that seed phrases and access keys could end up in the public domain due to a simple technical error. I recommend that users of AI assistants carefully check their publication settings and never share sensitive data through "Share" features without additional verification.