Conversations between users and AI chatbots surfaced unexpectedly in web search results over the weekend, revealing that private interactions can easily appear online through search engines such as Google and Bing. The issue, which was initially flagged by a user on Reddit, exposed shared Claude chats containing political advice inquiries, professional ethics questions for attorneys in Kansas, and erotic role play.
Anthropic Claude Chats Exposed in Search Engine Results
The exposure stems from how generative AI share features interact with web scrapers. Anthropic allows users to generate public URLs to share specific chatbot threads with others. While the company instructs web crawlers to avoid indexing shared chats using a robots.txt file at claude.ai that disallows the `/share/*` path, search engines treat these directives differently depending on page-level tags.
Technical guidance from search engines indicates that blocking a URL via robots.txt does not prevent it from appearing in results if other pages link to the target and a specific “noindex” HTML tag or response header is absent. WIRED reviewed exposed chat pages and confirmed they lacked the noindex tag. Google spokesperson WIRED stated that indexing shared chats is the responsibility of the platform rather than search engines, noting that search engines respect clear site-owner controls. Anthropic quickly deployed fixes and advised users to manage or unshare links directly through their account privacy settings, according to financialexpress.com.
Claude Mythos Uncovers Cryptanalytic Vulnerabilities
In a separate technical development, Anthropic published research revealing that its unreleased AI model, Claude Mythos Preview, discovered mathematical weaknesses in encrypted systems. According to Indiatimes, the model attacked a reduced 7-round variant of AES-128—which normally runs 10 rounds—using a technique 200 to 800 times faster than previous methods. The cipher normally underlies online bank transactions, encrypted chats, and Wi-Fi networks.
The model also halved the effective key strength of HAWK, a post-quantum signature scheme shortlisted by the US National Institute of Standards and Technology, after spending 60 hours on the task. Anthropic shared these findings with US government and industry partners prior to publication. Cryptographers routinely weaken algorithms experimentally to study patterns, and the AES attack assumed a high volume of chosen encryptions that do not occur on existing machines.
Discovery Process and Model Limitations
The discovery process involved unusual agent behavior. Initially, Claude declined to work on AES because it concluded no improvements were possible. After a researcher prompted the model to bypass its assumption that the task was impossible, Claude rewrote its agent harness to search for novel approaches.
Over three days, the system generated hundreds of millions of tokens, eventually producing roughly a billion tokens and landing on a fingerprinting technique named the Möbius Bridge. Two Anthropic researchers spent nearly a month verifying the mathematics. Anthropic also released CryptanalysisBench alongside ETH Zurich, Tel Aviv University, and the University of Haifa to track the evolution of these cryptanalytic capabilities.
Claude Mythos remains restricted due to safety precautions. When launched in April, Anthropic limited model access to select government agencies after determining it was unusually skilled at identifying and exploiting software bugs.
Security Comparison of Major AI Chat Platforms
| Platform | Reported Incident | Mitigation Status |
|---|---|---|
| Anthropic Claude | Shared chat URLs indexed in search engine results due to conflicting crawl rules. | Fixes deployed; users advised to manage share links via privacy settings. |
| OpenAI ChatGPT | Shared chat threads appeared in search engine results. | Pulled shared chats from search indexing in August 2025. |
| Google Bard | Transcripts surfaced via search indexing. | Blocked transcript indexing in 2023. |
