AI 周报：Claude Duet 上线、GPT-5.6 挖漏洞
Anthropic 推出主动式助理 Claude Duet（美国 Pro/Max 候补中），另有 GPT-5.6 找出 20 年 PHP 漏洞、决策模型、DeepSeek V4.1 Flash 等。
AI news roundup: • Anthropic launched Claude Duet — a proactive assistant that monitors your tools and suggests actions before you ask (similar to OpenAI’s Dots and xAI’s Grok Bots). Still waitlisted for US Pro/Max users. • OpenAI used GPT-5.6 to find a 20-year-old PHP vulnerability and released a free scanner so sites can patch it. Shows how AI is speeding up both discovery and exploitation of old bugs. • Decision models are getting real traction. OpenAI and Perplexity both shipped versions that output only a decision (yes/no or class choice) — much faster and cheaper than full LLMs. • Microsoft published MAI-DxO, claiming it outperforms doctors on medical diagnosis in their tests. • DeepSeek dropped V4.1 Flash (open-source, strong performance, very cheap). Tencent released Hy-Nomas, a high-quality world model for real-time virtual environments. Infrastructure and tools are moving fast. Most of it is still early.