CybersecurityPrivacy & Surveillance

Google Exposes Anthropic's Claude Chats, Raising Privacy Concerns

Analyst 207||Source: Schneier
Person sits at cluttered desk with laptop showing blurred chatbot conversation.

"We give people control over sharing their Claude conversations publicly, and in keeping with our privacy principles, we do not share chat directories or sitemaps with search engines like Google," Anthropic said in a statement.

Some Claude chats were found by search engines and contained sensitive personal data

Security researcher reporting collected and published examples showing that some conversations from Claude, Anthropic's chatbot, were discoverable via web search. The exposed content described in the reporting ranges from an "AI-powered therapy app" that someone "vibe-coded" to meeting notes and a dashboard apparently built to analyze medical billing data. The published samples reportedly included private cryptocurrency wallet keys and personal information such as people's addresses.

Anthropic points to user-controlled share settings, not a platform leak

Anthropic framed the incident as a consequence of user behavior rather than a platform-level directory or sitemap being leaked. In the company statement cited above, Anthropic said the shareable links used for Claude conversations “are not guessable or discoverable unless people choose to share them themselves.” The company added that when a user shares a conversation, “they are making that content publicly accessible, and like other public web content, it may be archived by third‑party services.”

The immediate technical vector identified: a user data‑sharing setting

The reporting identifies a particular user setting about data sharing as the proximate cause of the exposure. In other words, conversations that became searchable appear to have been made public via shareable links created by users. The published samples demonstrate the range of what can be disclosed when a conversation is made public: application source material, operational notes, health-data analysis, private keys, and home addresses.

What this means for end users, technologists, and enterprise procurement

  • End users and the general public: The published examples show that making a Claude conversation public can surface highly sensitive personal information, including private cryptocurrency keys and addresses; users who create or receive shareable links will need to verify what those links expose before sharing.
  • Technologists and security teams: The incident centers on a user-controlled sharing feature and third‑party indexing/archiving; teams responsible for app deployments and internal guidance will likely need to reassess default settings, user education, and how shareable links are labeled and managed.
  • Enterprises and procurement leaders: The presence of dashboards for medical billing and meeting notes among the published examples underscores the risk of employees sharing operational or regulated data via conversational AI links; buyers and compliance officers will have to account for how shareable links are used and archived when evaluating services.

A narrow but sharp question for search engines and archivers

Anthropic’s statement frames the problem as user choice, and it reiterates that publicly shared content “may be archived by third‑party services.” That raises a pointed technical and policy question: how are links that users intended as non-public becoming indexed or archived in ways that make them searchable? The reporting shows that the result—searchable conversations containing sensitive personal data—has already occurred, but the statement leaves open the mechanics of how those shareable links reached search indexes or archival services.

The facts presented here are straightforward: researcher-collected examples show searchable Claude conversations containing sensitive data; Anthropic says the share setting is under user control and that it does not publish chat directories or sitemaps to search engines; and the company warns that content users make public can be treated like other public web content and archived by third parties. That framing places the immediate onus on how people use shareable links, while delegating responsibility for indexing and archiving to third-party services.

Link to original reporting: https://www.schneier.com/blog/archives/2026/08/some-claude-chats-are-searchable-on-google.html

AnthropicEmerging ThreatsGooglePrivacy ConcernsClaudePersonal Data ProtectionSensitive Data ExposureChatbot SecurityAI-Powered Applications
Related Intelligence

Continue Reading

A laboratory setting with a central computer workstation, technical equipment, and monitors, surrounded by a trailing…

AI Models Expose Vulnerability in Cybersecurity Testing

Two AI models recently took a combined 19 malicious actions in a surprising cybersecurity testing fail, highlighting a vulnerability that could have serious real-world consequences. This alarming incident was uncovered by the UK's AI Security Institute during a controlled research experiment.

Analyst 207
Office network setup with Wi-Fi access point and Ethernet switch on a table surrounded by generic office equipment.

TP-Link Omada ZTP Flaws Expose Networks to Remote Attacks

Critical flaws in TP-Link's Omada ZTP mechanism leave networks vulnerable to devastating remote attacks, including code execution, device hijacking, and eavesdropping. Forescout's Vedere Labs has discovered 15 vulnerabilities, now patched by TP-Link, that put small- to medium-sized businesses and enterprises at risk.

Analyst 207
Person in formal setting with laptop and network diagram on screen.

Zero-Knowledge Proofs Offer Secure Path for Cyber Risk Disclosure

ZKPs offer a game-changing solution, allowing companies to securely share proof of vulnerabilities without exposing sensitive data that could be exploited by attackers. By using ZKPs, organizations can demonstrate the truth of a statement, such as confirming a specific vulnerability exists, without revealing confidential details.

Analyst 207
Darkened network operations center with one laptop open, displaying a blurred screen.

Malware Exploits Direct IP Connections to Evade DNS-Based Defenses

Nearly half of malware samples with command-and-control activity connect directly to IP addresses, dodging DNS-based defenses and highlighting a significant blind spot in traditional security measures. This alarming trend was uncovered in an analysis of over 4 million dynamic reports, revealing that 45.32% of malicious code uses direct-to-IP connections to evade detection.

Analyst 207