AI · Nov 12, 2026
Microsoft Research finds 8% of coding-agent sessions drive 44% of AI costs, plus a KLS proofNadella's uncomfortable admission: AI finds bugs faster than the world can safely fix them
· Updated · Signal Desk newsroom
Analysis of reporting by TechCrunch. How we work

At DealBook, Microsoft's CEO pointed to internet-connected firmware, chillers and HVAC systems that cannot be patched, and warned that instant patching under active exploitation creates its own instability.
Key takeaways
- Satya Nadella said much of critical infrastructure, including firmware, chillers and HVAC systems, is unpatchable yet connected to the internet.
- He said transportation systems also contain discoverable vulnerabilities in firmware that cannot be updated.
- AI has made vulnerability discovery powerful, he argued, but patches now must go straight to production when flaws are exploited, which creates instability.
- He described Microsoft's recent Outlook outage as a latent bug that led to a perfect storm, with architectural and operational fixes to come.
- Officials have denied that a wave of recent outages were cyberattacks.
The unpatchable layer
Speaking at The New York Times DealBook Summit, Satya Nadella offered a blunt description of the attack surface. A lot of critical infrastructure is unpatchable, he said, naming firmware, chillers, the cooling equipment common in data centers, and HVAC systems, and all of it is on the internet. He added that transportation systems contain discoverable vulnerabilities in firmware that cannot be updated.
That is not a new insight to security professionals, but it lands differently from the head of one of the world's largest software and cloud companies, and at a moment when AI-assisted attacks are a mainstream fear. Years of connecting ordinary equipment to networks created a vast fleet of devices that were never designed to be updated. In an era of cheap automated vulnerability discovery, those devices are the soft underbelly.
The patching paradox
Nadella's sharper point concerned speed. Organizations now have, in his words, a very powerful tool that finds vulnerabilities. Historically they had time to test a patch, understand the situation and roll it out. When a flaw is being exploited in the wild, the patch has to go from available to production instantly, and he warned that this will create instability.
This is the right frame and an underappreciated one. The debate about AI and security often fixates on attackers getting smarter. The quieter problem is that defenders are getting flooded with findings and racing to ship fixes, and every rushed change is a chance to break something. WIRED's hosts made a similar argument about last week's cluster of failures in their look at the outage wave. Recent outages at the FAA, the NSA, Apple and Microsoft were officially not cyberattacks, but they show how fragile complex systems are under constant change.
Microsoft's own glass house
Nadella also addressed Microsoft's Outlook disruption, calling it a latent bug that produced a perfect storm and promising upstream architectural and operational fixes. That candor is welcome, though Microsoft is far from a neutral commentator: it sells security products, cloud infrastructure and AI tools that both find and fix vulnerabilities. His call to modernize, assess the threat and improve real-time detection and defense describes a market Microsoft would like to serve.
It also fits a pattern in his recent remarks. In October he argued that all AI models should be assumed compromised and that powerful models need an external emergency brake. The common thread is containment over prediction: assume failure and limit the blast radius. For infrastructure that cannot be patched, that means network isolation and monitoring rather than hoping a fix arrives. The weakness in his framing is cost. Replacing legacy controllers in buildings, plants and transit systems is slow and expensive, and nothing he said suggests who pays.
Who feels it
- Critical infrastructure operators
- Unpatchable devices need segmentation and monitoring, plus replacement budgets.
- Security teams
- AI-driven discovery increases the volume and urgency of patches to test and ship.
- Data center operators
- Cooling and building systems are part of the attack surface, not just servers.
- Policymakers
- Modernization of legacy industrial systems may need public funding or mandates.
What to watch
- Microsoft's detailed postmortem and architectural changes after the Outlook outage.
- Federal moves to inventory or fund replacement of unpatchable industrial devices.
- Industry guidance for testing emergency patches without full production risk.
- Evidence of AI-assisted exploitation of legacy firmware in the wild.
FAQ
What did Satya Nadella say about critical infrastructure?
He said a lot of critical infrastructure, such as firmware, chillers and HVAC systems, is unpatchable but connected to the internet, and that transportation systems also have unpatchable firmware with discoverable vulnerabilities.
Why does Nadella think AI creates instability?
AI makes finding vulnerabilities much faster, so patches for actively exploited flaws must go straight to production with little testing time.
What caused Microsoft's Outlook outage?
Nadella described it as a latent bug that led to a perfect storm and said Microsoft will make architectural and operational fixes.
Companies: Microsoft
More on Microsoft
AI · Oct 29, 2026
Microsoft research finds coding agents struggle most with reading unfamiliar codeAI · Oct 30, 2026
A single robots.txt line to block AI training: what the OpenAI-Google-Microsoft-Anthropic pledge does and doesn’t fixAI · Nov 19, 2026
At Ignite, Nvidia and Microsoft pitch one stack from Vera Rubin racks to RTX Spark laptopsAI · Nov 12, 2026
Jensen Huang’s exit gives Nvidia its first board chair, and that governance change is the real story