“Really, really nervous”: Microsoft’s AI CEO says AI improving AI is a dangerous capability, but Microsoft hasn’t hit the brakes as it puts AI agents on your Windows 11 PC
windowslatest.com Nov 22, 2026

“Really, really nervous”: Microsoft’s AI CEO says AI improving AI is a dangerous capability, but Microsoft hasn’t hit the brakes as it puts AI agents on your Windows 11 PC

AI-summarised brief · reviewed before publication

Microsoft’s AI chief Mustafa Suleyman warned that AI models are now “fully involved” in improving the next generation of models, describing the recursive self‑improvement (RSI) capability as “definitely a dangerous capability.” Speaking to former CNBC correspondent Deirdre Bosa on November 17, the day Microsoft unveiled six new MAI models at its Ignite conference, Suleyman said the past three months marked a “total watershed moment” as internal models begin to debug, generate training data and evaluate themselves. While engineers remain in the loop, the company is accelerating the deployment of frontier‑level models, including MAI Code 2, which it claims matches 98 % of GPT‑6.1 performance at 40 % of the cost, and plans to embed AI agents in Windows 11 and Microsoft 365. The push aims to outpace rivals OpenAI and Anthropic, even as internal staff voice concerns about speed‑over‑safety culture.

💡 Why It Matters

  • · Microsoft is betting its core productivity ecosystem on self‑enhancing AI, turning the very technology it warns about into a competitive differentiator.