Anthropic's Shared Claude Chats Leaked to Google and Bing

Anthropic's AI assistant Claude includes a "Share Chat" feature that lets users share conversations. It has now emerged that URLs generated by this feature were inadvertently indexed in Google and Bing search results, leaving conversations that should have remained private viewable by anyone.

According to reports from TechCrunch and WIRED, the root of the problem is believed to lie in how the shared URLs handled crawler control. The design itself — anyone who knows the URL can access the conversation — was intentional, but search engines indexing those URLs was reportedly an unforeseen consequence.

This issue serves as an important warning for users who tend to assume their conversations with AI chatbots are "private." When entering confidential or personal information into an AI chat, it's worth being cautious regardless of whether a sharing feature exists.

Moonshot AI Releases Kimi K3 Open Weights

China's Moonshot AI has open-sourced the model weights of its latest model, Kimi K3, along with parts of its infrastructure setup. According to The Decoder, Kimi K3 shows performance approaching Western frontier models such as GPT-5.6 Sol and Fable 5 on common benchmarks.

However, independent evaluations have found significant gaps in the fields of cybersecurity and mathematics, and the possibility of distillation from other companies' models has also been raised. A Bloomberg opinion piece even describes the event as "a Sputnik moment for OpenAI."

Over on Reddit's LocalLLaMA community, there are also reports of attempts to analyze Kimi K3's 896-expert structure using the HuggingFace Viewer, showing how the open-weights release is energizing community analysis.

Nvidia Invests $5 Billion in Ilya Sutskever's SSI

According to a Financial Times report (via Hacker News), Nvidia is said to be investing roughly $5 billion in SSI (Safe Superintelligence Inc.), the safe superintelligence company founded by OpenAI co-founder Ilya Sutskever. The move is drawing attention as a sign that AI infrastructure companies are getting serious about backing the development of next-generation foundation models.

OpenAI Report: 43.5% of ChatGPT Users Are Doing "Other People's Jobs"

OpenAI analyzed more than 800,000 work-related ChatGPT messages and found that 43.5% of occupation-specific questions relate to tasks belonging to other occupations. The company calls this phenomenon "task crossover."

The trend is especially pronounced at small and medium-sized businesses, revealing that the fewer specialists an organization has, the more it compensates with ChatGPT. It's an interesting piece of data suggesting that AI tools are gradually blurring the boundaries between roles within organizations.

Clean Code Author Robert Martin: His "Don't Read It" Strategy for AI-Generated Code

Robert C. Martin (Uncle Bob), author of the software engineering classic Clean Code, shared an interesting perspective on his X (formerly Twitter) account. Martin, who started coding in the late 1960s, says he now follows a strategy of never reading the code his AI agents write.

Instead, he surrounds his agents with "extreme constraints" — unit tests, Gherkin tests, quality metrics, mutation testing, and coverage measurement — and accepts only code that passes the tests. On Hacker News, this approach sparked both praise and criticism for a strategy that some call the ultimate form of test-driven development.

Professor's "Invisible Prompt Trap" Catches 32 of 35 Students Using AI

A university professor devised a method to detect whether students were using AI to cheat on assignments: embedding an "invisible prompt," imperceptible to the human eye, into the assignment instructions. The result: 32 out of 35 students were found to have used AI.

The case drew considerable attention, gathering 30 points on Hacker News and prompting active debate about where the line lies between surveillance and ethics in educational settings when it comes to AI use.

---