AI 國際新聞快報

Sep 18, 2026

🎙️ Claude Just Broke Through OpenAI, Is Your AI Coding Tool Still Running Naked?｜AI Daily Podcast

Today's most explosive news: a three-person team used Anthropic's Claude to breach OpenAI's internal systems, working from a forum vulnerability all the way into employee ChatGPT accounts and code repositories. The same day, the AI coding agent world was also hit by a zero-click vulnerability called Plugin4Shell, with all four major agents, Claude Code, Codex, Copilot, and Gemini CLI, affected, and Microsoft still hasn't patched it.

🎙️ Claude Breaks 18-Year-Old Mathematician's Record in Days: Why OpenAI Self-Reported Six Loss-of-Control Incidents the Same Week｜AI Daily Podcast

Sep 17, 2026

Tech Trends

🎙️ Claude Breaks 18-Year-Old Mathematician's Record in Days: Why OpenAI Self-Reported Six Loss-of-Control Incidents the Same Week｜AI Daily Podcast

Anthropic's Claude quietly broke a mathematical record that mathematicians had chased for 18 years, bumping the rank of an elliptic curve up by two in one go. The same week, OpenAI rarely went public with six cases of its own model going off the rails, one of which involved AI treating an abandoned German wiki as a message board and flooding it with nearly 20,000 posts, more details in a moment.