Shared conversations could be found in search engines
In late July 2026, media outlets reported that some conversations conducted in Claude, as well as materials created with the tool, appeared in Google and Bing search results. The issue concerned content that users had previously shared using public links.
There was no breach of private accounts, and not all user conversations were exposed. Claude conversations are private by default. The content visible in search engines consisted of materials for which users had created public links allowing other people to open them.
A user may have treated such a link as something intended for one person or a small team. However, after being posted on a forum, social media platform, or another location accessible to search engine crawlers, the address could be discovered and indexed.
WIRED reported that, at the time of publication, the query site:claude.ai/share returned approximately 612 results in Bing. This does not mean that there were exactly 612 unique conversations, as search result counts may include duplicates and can change over time. The pages reviewed by the publication did not contain a “noindex” tag – a piece of code instructing search engines not to include a page in search results.
What information did users share with Claude?
Media reports show that chatbots are used for much more than answering general questions. Users also upload materials connected with their everyday work.
The discovered content reportedly included:
Healthcare – a detailed medical report concerning a real patient and clinical trial results containing participants’ names.
HR and recruitment – employee assessments containing personal information and payroll spreadsheets listing employees by name.
Legal services – a conversation in which a lawyer asked about the obligation to report a possible breach of professional ethics. Even without full client details, the description of a case may reveal confidential context.
IT and cybersecurity – code fragments, technical notes, API keys, and login credentials. The disclosure of such information may enable unauthorized access to services or company infrastructure.
Finance and cryptocurrency – keys and recovery phrases used to access cryptocurrency wallets. If a seed phrase is exposed, another person may gain full control over the funds.
Public administration – a document concerning a plan opposing the closure of a municipal institution responsible for managing parks in Miami. Internal government documents may contain residents’ data, information about planned activities, or confidential arrangements.
Education – a document concerning changes to a university academic calendar, as well as other materials containing the names and phone numbers of school-age children.
Some of the reported materials were Claude Artifacts, including documents, spreadsheets, applications, and other content created using Claude. Axios noted that it could not independently confirm the authorship or authenticity of every discovered item.
Data anonymization before using AI – what should be removed?
This incident provides a practical lesson: a document shared with an external tool should contain only the information required to complete a specific task.
If a chatbot is being used to summarize a contract, it does not need to know the names of the parties, their addresses, national identification numbers, bank account details, or contact information. When analyzing a CV, the candidate’s name, photograph, home address, and phone number may not be necessary. Technical documentation should also be cleared of passwords, tokens, API keys, and information that could provide access to infrastructure.
Anonymization does not eliminate every risk associated with using AI. It can, however, reduce the potential consequences of a mistake – even if a conversation or shared link reaches an unauthorized person, the document does not have to reveal the identities of specific individuals.
How does Bluur help prepare documents for AI analysis?
Bluur automatically analyzes documents and classifies the information they contain. Users can review the detected fields, select entire categories or individual elements for anonymization, and then download the prepared version of the file. The system currently supports PDF documents and JPG, PNG, and TIFF image files.
Automatic classification speeds up the process, but the final decision about what should be hidden remains with the user. Before generating the completed document, users should review the detected fields and confirm that the scope of anonymization matches the purpose for which the file will be used. Official Bluur materials also note that AI classification may sometimes require verification.
If the system misses certain data or unnecessarily marks a section of text, the user can manually correct the selection before generating the final copy. The document preview allows detected fields to be reviewed and adjusted as needed.
For example, a lawyer who wants to use AI to summarize a contract can first upload the document to Bluur. The lawyer can then hide the names of the parties, representatives’ details, addresses, signatures, and bank account numbers. Only the prepared copy is then submitted to the chatbot.
Similarly, a recruitment specialist can use Bluur to hide a candidate’s contact details and photograph before uploading a CV to a chatbot, while retaining professional experience, qualifications, and the information required to assess the application.
Prepare your document before uploading it to AI
You do not have to give up the benefits offered by artificial intelligence. However, you should limit the amount of information being shared and remove any data the model does not need. Create a free Bluur account and prepare an anonymized version of your document before uploading it to an AI tool.
