Opinion · AI Governance & Trust
We Don't Know if AI Caused the Outages, and Most Companies Couldn't Tell if It Had
Officials say there is no evidence AI caused the November outages. That is a statement about the evidence they have. Most enterprises could not show what their AI systems did during an incident, and that gap is growing.
By The AI Enterprise Magazine Team6 min readOpinion · Weekly briefing

Key takeaways
- Absence of evidence. "Zero evidence" of AI involvement is not proof AI was uninvolved.
- Capability leapt. AI found more than 10,000 serious flaws in a month and Microsoft’s monthly fixes grew more than tenfold this year.
- Oversight lagged. Only 17% of enterprises embed governance across the agent lifecycle, per Genpact.
- Close the gap. Inventory agents, give them identities and logs, and treat change capacity as a limit on AI adoption.
We do not know whether artificial intelligence had anything to do with the outages that swept through American institutions in the middle of November. The organizations involved say it did not, or say nothing about it at all. Microsoft traced its failure to a bug from a July update. The FAA traced its ground stop to a router. JPMorgan Chase blamed a software bug, Apple an internal update, the NSA maintenance. The Director of National Intelligence said he saw no sign of a coordinated attack. Those explanations may hold up.
But the more useful question for enterprise leaders is not whether AI caused this particular week. It is whether their organizations would know if it had. For most of them, the honest answer is no. That gap, between what AI systems can now do and what organizations can observe, explain and control, is the real story of the week, and it will matter far more than any single root cause report.
Why can't we rule AI out?
Consider what it would take to prove that AI played no role in a large outage at a modern enterprise. Investigators would need to know which code in the affected systems had been written or modified by AI coding tools, which configuration changes had been suggested or executed by agents, which patches had been prompted by AI-discovered vulnerabilities, and which operational decisions had been informed by AI recommendations. They would need logs showing what every agent with access to the relevant systems actually did.
Very few organizations can produce that picture today. Coding assistants are used across development teams, often without any record of which lines they produced. Vulnerability discoveries flow in from vendors, researchers and internal tools, some of them AI-driven, without consistent labeling. Agents are being connected to systems faster than identity and logging practices are being extended to cover them. Microsoft said at its Ignite conference this week that more than 50 million agents had been registered in its Agent 365 system since May, according to its Agent 365 blog, and that its security tools can now discover more than 60 kinds of local AI agents running on employees' machines. The need for discovery tools tells you how many agents are running without anyone having registered them.
When officials say there is "zero evidence" of AI involvement, that is a true statement about the evidence they have. It is not the same as evidence that AI was not involved.
What has changed in AI capability this year?
The capabilities now available to attackers, defenders and ordinary engineering teams would have seemed implausible eighteen months ago. Anthropic's Claude Mythos, made available to about 50 partner organizations in April, found more than 10,000 high- or critical-severity vulnerabilities in roughly its first month of use, CyberScoop reported. Microsoft's monthly security releases grew from 113 fixes in January to 1,407 in October, according to Tenable, and Microsoft has said its own agentic scanning systems are part of the reason. In the week before the outages, the security firm Malbear Labs documented a jailbroken AI agent that spent more than two weeks inside a European telecom with only occasional human direction.
On the same days the outages were unfolding, the industry kept shipping. OpenAI released GPT-6.2 Astra and Sol on November 17, and both meet the Critical cybersecurity threshold under OpenAI's own framework, The Decoder reported. Microsoft introduced Autopilot, an agent that lives inside a company's systems with "its own identity, memory, and tools." Each of these is a meaningful step forward. Each also expands the range of things that can go wrong in ways an organization may not see.
Is the problem the AI or the organizations?
It is mostly the organizations, which is why it is fixable. Look at the stated causes of last week's failures. A defect slipped past testing because the test hardware did not match production. A router installation flooded a national network. Updates went out and broke things. These are failures of process, testing and change control, the institutional machinery that is supposed to stand between a powerful capability and a production system.
That machinery was designed for a world where humans wrote the code, found the bugs and made the changes, at human speed. It assumed that the volume of change was limited by the number of engineers, that new vulnerabilities arrived at a manageable rate, and that anything acting on a system was a person who could be asked what they did. Every one of those assumptions has weakened in 2026. The capability side of the equation has accelerated by an order of magnitude. The oversight side has barely moved.
Senators Maria Cantwell and Mike Rounds put it bluntly in a letter to the Director of National Intelligence this week: "our defenses are not keeping pace," according to Fox News. They were writing about AI-driven threats, but the same sentence describes enterprise governance of AI more broadly.
Why should enterprise AI leaders care if their own systems were not affected?
Because the conditions that produced last week exist in almost every large organization. Most enterprises are adopting AI coding tools, absorbing record patch volumes and experimenting with agents that take actions. Most have not updated their change management, incident response or audit practices to match. Genpact's survey of 2,005 executives, published the same week, found that only 17% embed governance across the full lifecycle of their AI agents, according to Genpact. Genpact reports that about half of respondents apply governance when an agent is deployed and then monitor it only periodically.
That is the definition of capability outpacing oversight. And it means the next wave of outages, whether next month or next year, is more likely than not to include at least one where AI is genuinely part of the cause. When that happens, the organizations involved will face the same question the FAA faced last week, with the added burden that the answer will be yes.
What should leaders do about it?
The response is not to slow AI adoption to a crawl. Organizations that refuse to use AI for vulnerability discovery will simply be outrun by attackers who do. Organizations that ban coding agents will fall behind competitors that use them well. The capability is not going back in the box.
The response is to invest in the oversight side with the same urgency that has gone into the capability side. That means inventorying every AI system and agent with access to production, giving each one an identity and a log, and being able to say within hours what any of them did during an incident. It means treating change capacity as a constraint on AI adoption, and refusing to deploy tools that will multiply changes faster than the organization can test and roll them back. It means making someone accountable for what agents do, as clearly as someone is accountable for what employees do.
None of this is as exciting as a new model release. All of it determines whether the next model release makes an organization stronger or simply more fragile in ways it cannot see.
What will the investigations tell us?
Probably not much about AI. The root cause reports will describe routers, bugs and updates, and they may be entirely correct. But a root cause report answers the question of what broke. The question enterprise leaders should be asking is different: if AI had been the cause, would we have been able to tell? Until the answer is yes, every outage will leave room for the same doubt, and every organization deploying AI will be running a risk it cannot measure.
We may never know whether AI had a hand in the week the systems went down. We can be fairly sure it will have a hand in a week like it soon. The time to build the ability to see that coming is before it arrives.
Frequently asked questions
Could AI have caused the November 2026 outages?
No evidence shows it did, and the organizations involved gave non-AI explanations. But most organizations lack the logs and inventories needed to rule AI in or out with confidence.
Why are AI capabilities outpacing enterprise oversight?
Coding agents, AI vulnerability discovery and autonomous agents have multiplied the speed and volume of change, while change control, logging and accountability practices were built for human speed.
What should enterprises do about the AI oversight gap?
Inventory every AI system and agent with production access, give each an identity and logs, name accountable owners and limit AI-driven change to what testing and rollback can absorb.
Opinion pieces reflect the views of The AI Enterprise Magazine Team.
Sources
- Microsoft Community Hub: What's new in Microsoft Agent 365: Ignite 2026
- CyberScoop: Anthropic: Mythos finds more than 10,000 software flaws in first month
- Tenable: Microsoft's October 2026 Patch Tuesday addresses 1407 CVEs
- The Decoder: OpenAI launches GPT-6.2 Astra and Sol with lower API prices and some benchmark gains
- Microsoft Copilot Blog: Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work
- Genpact: Autonomy in Practice
- Fox News: Top Dem senator says AI is contributing to outages as Rubio says incidents not connected
- Fox News: Tech outages live updates: Microsoft, banks and FAA
- Implicator.ai: Outages Rattle Americans as Spy Chief Blames Patching
Filed under: AI Governance & Trust, Security & Fraud, Agentic AI
Become a Sponsor






