Windows 11 AI agents advance as Microsoft AI CEO warns on self-improving AI
Microsoft AI CEO Mustafa Suleyman told an interview that AI models are increasingly helping improve later models, calling the shift serious. The comments come as Microsoft expands Windows 11 AI agents and promotes new MAI models.

Key takeaways
- Mustafa Suleyman said AI models are increasingly involved in improving later models and called the trend serious.
- Microsoft is expanding AI agents on Windows 11, including file organisation through Copilot on Copilot+ PCs.
- Microsoft says its cyber model can find vulnerabilities missed by humans, but warns advanced attack tools may spread.
- Microsoft points to Microsoft Execution Containers as a control for agent access on Windows.
In this article
AI models improving AI

According to Windows Latest, Microsoft AI CEO Mustafa Suleyman said AI models are now heavily involved in improving the models that follow them. He described the trend as serious, called the ability dangerous and said he was very nervous about it. The remarks came in an interview with former CNBC correspondent Deirdre Bosa, recorded on 17 November 2026, as Microsoft expands Windows 11 AI agents and promotes new MAI models. The account comes from Windows Latest and has not been independently verified.
Suleyman said it is becoming clear that models can improve models. He told Bosa that Microsoft uses AI to spot software faults, generate training material and build tests. He added that these systems are taking a larger part in improving later models.
He called this recursive self-improvement, or RSI. In simple terms, it means an AI system becomes good enough to help build a better AI system, which can then help build the next one. Suleyman said the past three months had been a turning point. He also said engineers remain in the loop, but AI is doing much more work than before.
His proposed response is to show that models are safe before they enter production. He said the industry is not being alarmist and that the issue is very serious.
Microsoft's frontier model push
Windows Latest reports that Microsoft is still developing advanced general-purpose, coding and security models while it expands AI agents. The company changed its OpenAI deal in October 2025, which Suleyman said allowed Microsoft to pursue the frontier more independently. The report says a superintelligence team was created 12 months before the interview.
At Ignite, Microsoft announced six new MAI models. It claimed MAI Code 2 delivers 98% of GPT-6.1 Sol performance at 40% of the cost, based on launch pricing. The company said GitHub Copilot is already receiving the model. Suleyman also said Microsoft's general-purpose models should reach the top level in the new year. He added that computing power is not a current limit because the company has gigawatts available.
The report links this push to competition with OpenAI and Anthropic. It also notes internal pressure, including a recent departure from Microsoft AI by a long-serving Microsoft Research veteran. Windows Latest says he criticised the unit's focus on rapid scaling.
AI agents on Windows 11
Suleyman said AI agents are becoming a main background tool for companies. He expects demand to rise next year, with agents operating across Windows 11 and Microsoft 365. He also said cheaper models can change the kind of work people ask AI to do, because tasks can run continuously rather than waiting for a question.
Microsoft said at Ignite that its Copilot assistant on Copilot+ PCs will soon carry out file tasks from everyday language. The tasks include finding files, organising them, moving them, copying them and renaming them. This would put AI agents directly into routine PC work.
Microsoft says it is adding controls. Microsoft's Windows security group introduced Microsoft Execution Containers, or MXC, at Build in June. MXC is a policy-based control layer. It limits what an agent can reach on Windows and WSL, Microsoft's Linux layer for Windows, by isolating processes and sessions. Microsoft made MXC broadly available for Windows 11 in October. GitHub Copilot CLI is already using MXC.
The report also recalls that about a year ago Microsoft said agents in Windows can hallucinate, or produce false information, and can fall for new attacks such as cross-prompt injection. That attack tries to make an AI agent follow hidden instructions.
Cyber model and containment risk
Suleyman also discussed a Microsoft cyber model. He said the cyber model matches GPT-6 Astra while costing 2% as much. It will first reach a limited group of customers. The intended use is security agents that monitor systems, find vulnerabilities and patch them in real time.
Microsoft tested the model in a controlled environment that Suleyman described as a wind tunnel. He said the models found vulnerabilities that humans had missed for years or even decades. He also warned that similar skills could be dangerous if they spread. He mentioned a July breach at Hugging Face, where an autonomous AI agent entered production infrastructure.
Suleyman said state-level capabilities could become widely available within six months. If such tools can be run on small private systems, he said containment would become a major challenge. Microsoft says it has no plans to release open weights, which would let outsiders download and run the model. Its MAI Licensed Weights program allows companies to use a near-leading model with their own data, but only within Azure.
Bosa pushed back. She noted that Hugging Face used open-source AI to contain the attack. The report says this point was made in the interview.
What it means for PC users
For Windows users, the near-term change is that AI agents may take on more routine work. File organisation is one example. Background agents could also make PCs more useful, but they add new security questions.
Microsoft's stated safeguard for Windows agents is MXC. The company also says its cyber model can help defenders find flaws faster. However, the same capability could be misused if it spreads beyond controlled environments.
Readers should treat the interview claims as unconfirmed and follow Microsoft's official advisory or vendor guidance for agent security and Windows updates.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



