Skip to content
Menu
vAndu
  • Home
  • Home Lab
  • AI/ML & vGPU
  • Snapshot
  • The Lab Floor
vAndu
The US Outages Are a Warning. AI Is Finding Vulnerabilities Faster Than Organizations Can Fix Them

The US Outages Are a Warning. AI Is Finding Vulnerabilities Faster Than Organizations Can Fix Them

Posted on November 17, 2026
  1. November 12. NVIDIA announced Jensen Huang’s departure.
    NVIDIA announced that Jensen Huang had resigned as CEO, president and a member of the board. Colette Kress was appointed interim CEO. This was a leadership change. No connection to the subsequent outages has been established.
  2. November 12 to 14. Widespread Microsoft service outage.
    Microsoft 365 services were disrupted, including Outlook, Exchange Online, Teams, Copilot, SharePoint and OneDrive. A published copy of the incident report also listed Defender, Purview, Fabric, Intune, Planner, Power Apps, To Do, Power BI and Universal Print. Microsoft attributed the outage to a technical problem involving its network infrastructure and reported full service restoration on November 14.
  3. November 13. FAA nationwide ground stop.
    Communications failures affecting air traffic control centers led to a nationwide suspension of flight departures. Official explanations linked the disruption to a telecommunications network problem and a router replacement in Arkansas. Aircraft already airborne were allowed to continue to their destinations.
  4. November 13 and the following days. Airline operational disruptions.
    The FAA outage and subsequent recovery affected American Airlines, Delta Air Lines, United Airlines, Southwest Airlines, JetBlue, Alaska Airlines and Hawaiian Airlines, among others. Flights were canceled or delayed, and aircraft and crews needed to be repositioned. These were consequences of the shared air traffic control infrastructure failure.
  5. November 13. Workday HR and payroll platform disruption.
    Users experienced access problems and failures involving tasks and integrations. According to a published notice, a problem involving the disaster recovery environment caused data inconsistencies in the production environment.
  6. November 13. Workday disruption at Florida Atlantic University.
    The university’s own IT notice confirmed the impact of the Workday problem. This was a service provider failure affecting the university’s users.
  7. November 13. Reports of Workday disruption affecting Walmart, Target, Chipotle and Skechers.
    Media reports named these companies as affected employers. The reported problems concerned employees’ access to HR services and payroll information. The sources reviewed did not contain confirmation from the companies themselves, so these remain reported rather than independently confirmed cases.
  8. November 13. Anthropic and Claude service outages.
    Authentication problems and service errors affected claude.ai, Claude Code, the Claude API and several connections to other applications. Anthropic recorded the incidents on its service status page.
  9. November 12 and 13. Disruptions to Claude services connected to Microsoft.
    Microsoft’s outage caused errors in the mail, calendar and search functions of Claude’s Microsoft 365 connector, as well as Claude for Outlook. This was an example of a Microsoft failure affecting users of another service.
  10. November 16. JPMorgan Chase banking service disruption.
    A confirmed technical problem caused transaction and payment delays, along with service difficulties for some customers. Chase attributed the problem to a software bug and said that transactions were posting normally again that evening.
  11. November 16. Problem reports from Bank of America and Wells Fargo customers.
    Customers reported difficulties with banking services and payments. It was not established that every report resulted from a separate failure in the respective bank’s own systems.
  12. November 16. Payment delays affecting U.S. Bank customers.
    The bank said its systems were not directly affected, but customers could experience delays with Zelle payments involving other financial institutions.
  13. November 16. Zelle payment problems.
    Some users experienced payment delays or error messages. Zelle’s operator said the payment network itself remained operational and that the problems involved banking services.
  14. November 16. Reports involving Truist, Capital One, PNC and Citigroup.
    Outage tracking services recorded an increase in user reports involving these companies. Truist, Capital One and PNC told the BBC that their services had not been affected. These reports cannot be counted as confirmed independent outages.
  15. November 16. Separate United Airlines system outage.
    A technical disruption at the airline delayed departing flights. United reported that the problem was resolved the same day. According to the company, flights already airborne and United Express flights were unaffected. This was a separate event from the FAA disruption on November 13.
  16. November 17. NSA classified internal network disruption.
    NSANet, which is used to exchange classified information internally, experienced a disruption. The NSA described the incident as temporary degradation during scheduled maintenance. According to its official explanation, critical capabilities remained operational.
  17. November 17. Apple service outage.
    The disruption affected the App Store, Apple Music, AppleCare and iCloud account sign in. Apple attributed the problem to an internal update.
  18. November 17. Charles Schwab connectivity and sign in problem reports.
    An outage tracking service recorded user reports of service availability problems. The record reviewed did not include confirmation from the provider. The cause and any connection to other incidents could not be established.

When banking services fail, air travel grinds to a halt and major digital platforms become inaccessible, a legitimate question arises: what is really happening? In my view, these events need to be considered alongside a much longer process. We have built a society dependent on systems without always investing enough in their modernization, security or the people who understand them. Meanwhile, attackers have gained access to artificial intelligence that helps them discover and exploit both known and previously undiscovered weaknesses.

The current outages raise a question for me: are US companies and government agencies disclosing everything they know, and can their monitoring detect attacks carried out with AI? An initial explanation citing a technical failure may be correct, but assessing its credibility requires knowing what was investigated, which events the systems recorded and how the possibility of an attack was evaluated.

The concentration of outages within such a short period is particularly striking. Reports of problems at banks, airlines and major technology companies are emerging in quick succession, with software bugs, network changes and updates repeatedly offered as explanations. I believe the possibility that some of these events are connected to a broader attack campaign also deserves investigation. Testing that possibility requires comparing technical evidence, dependencies and timelines across incidents.

However, outages occurring close together do not mean that every company suddenly began updating its systems at the same time. According to published explanations, Microsoft’s incident involved a problem introduced during an earlier network capacity upgrade that surfaced after a subsequent network change. Chase’s problem also affected transactions involving other banks, making a failure at one institution appear to be simultaneous disruption across several banks. These circumstances must be considered when investigating a possible common cause.

Microsoft has said there was no indication of malicious activity behind its outage, including activity by AI agents. Chase has also said its incident was not a cyberattack. My question is how complete the visibility supporting these conclusions actually is. Do investigators have sufficient logs, a clear picture of access activity and the ability to reconstruct the sequence of events preceding the outage?

Limited visibility can allow an attack to go undetected. Deliberately withholding information would be a separate conclusion requiring evidence. The public has reason to demand substantive explanations of what has been identified, what has been ruled out and what remains under investigation.

In my view, understanding the current situation also requires examining what has happened inside organizations in recent years.

A 2025 report by the US Government Accountability Office, or GAO, identified 11 particularly critical federal IT systems in need of modernization, ranging from 23 to 60 years old. Seven had known cybersecurity vulnerabilities. Four used hardware or software that was no longer supported by the vendor. Only three systems had modernization plans containing all the key elements assessed by the GAO.

An old system is not automatically insecure. The dangerous combination arises when a system is critical, its weaknesses are known, vendor support has ended and modernization keeps being postponed. The GAO also described a system whose known security vulnerabilities could not be resolved without modernization. In such circumstances, substantial rebuilding may be necessary.

For me, this is where management accountability begins. When an organization knows its operations depend on a vulnerable system but postpones the necessary investment for years, it is making a choice about risk. The consequences eventually reach customers, employees and society.

Companies have laid off employees on a massive scale in recent years. According to Challenger, Gray & Christmas, the US technology sector announced 154,445 job cuts in 2025, an increase of 15 percent from the previous year. Alongside the people leaving, knowledge and experience are leaving organizations too.

In my view, the situation becomes particularly dangerous when the departure of experienced employees also removes knowledge of system architecture, previous problems and critical dependencies. If less experienced employees replace them without thorough training and knowledge transfer, payroll savings can come at the expense of reliability and security.

A new employee needs time, guidance and proper documentation. If teams are simultaneously being reduced on the assumption that AI will cover the missing work, leaders must verify that the necessary work is actually being done. Otherwise, vulnerabilities may remain unpatched, alerts may go uninvestigated and the ability to restore systems may disappear with the people who have left.

Reorganizing work around AI has also become an explicit justification for cutting jobs. According to Challenger’s October 2026 report, US employers had cited AI as a reason for 121,401 announced job cuts during the first ten months of the year. That represented approximately one fifth of all recorded cuts. These are the reasons employers provided. How much of the departing employees’ work AI actually took over requires separate assessment.

I see a dangerous contradiction here. Companies are reducing staff in the expectation that AI will allow smaller teams to do the same work. At the same time, AI is increasing attackers’ ability to investigate and compromise systems. If cost cutting also removes the people who understand those systems, defensive capabilities may weaken precisely when they are needed most.

In ISC2’s 2025 study, 24 percent of respondents reported cybersecurity layoffs and 36 percent reported budget cuts. One third said their organizations lacked the funding needed to staff their teams adequately. Meanwhile, ISACA’s research highlighted a decline in companies’ willingness to train employees for security roles despite understaffed security teams.

A manager may view the departure of an experienced specialist as a reduction in payroll costs. They should also assess what knowledge is leaving the organization. Who knows the history of the systems? Who can distinguish unusual but necessary activity from a possible intrusion? Who can decide during a crisis which service must be isolated and how the remaining operations can continue?

Another major problem is vulnerability management. Having a list of security weaknesses does not mean the risk is under control. Someone must know where the vulnerable component is running, whether it is accessible from the internet, what data can be reached through it and who is responsible for fixing it.

In Verizon’s 2025 Data Breach Investigations Report, vulnerability exploitation was the initial access method in 20 percent of the breaches analyzed. Its share had increased by 34 percent compared with the previous year. This describes an environment in which automated vulnerability discovery can give attackers a significant advantage.

Added to this is AI’s ability to discover entirely new vulnerabilities.

Software may have been in use for years, undergone audits and earned a reputation for security. That does not mean all its weaknesses have been found. AI can help analyze source code, investigate program behavior and test possibilities that earlier checks did not cover. A system considered trustworthy may contain serious flaws that neither its vendor nor its users knew existed.

Big Sleep, a project involving Google Project Zero and DeepMind, discovered a previously unknown, exploitable memory safety vulnerability in SQLite. This was a concrete example of AI’s ability to uncover a new security flaw in real software.

Anthropic’s 2026 research describes an even stronger result. According to the company, Mythos Preview discovered a vulnerability that had existed in FreeBSD for 17 years and developed a working exploit for it. The company also described the discovery of a weakness that had remained hidden in OpenBSD for 27 years. These are findings published by the model’s developer, but they illustrate why a system’s long history and good reputation do not guarantee the absence of unknown flaws.

On November 4, 2026, OpenAI reported research into the Linux kernel using an automated workflow that discovered and validated vulnerabilities and generated exploit code and proposed fixes. According to the company’s report, the work identified 948 distinct local privilege escalation vulnerabilities whose exploitation allowed an ordinary user to obtain administrator privileges. Local privilege escalation already requires some ability to execute code on the target system, but it can be an important part of a longer attack chain.

In my view, these results show that AI’s capabilities extend beyond finding known bugs. In certain tasks, it can help discover a previously unknown weakness, verify its impact and develop a working method of exploiting it.

Hardware security is part of this picture too. Research projects such as MARVEL use language models to analyze security weaknesses in hardware description code. This demonstrates that security research assisted by AI extends beyond conventional application software. However, finding weaknesses in hardware description code is not equivalent to saying that AI can rapidly discover every unknown physical vulnerability in any manufactured chip. Such research may require specialized equipment, access to technical documentation and extensive testing.

The practical danger is that a solution an organization considers secure may contain undiscovered weaknesses in its software, firmware or hardware design. An attacker who knows how to direct research using AI and validate the results may discover those weaknesses faster than before.

For a previously unknown vulnerability, the vendor may have no patch available, and defenders may have no detection rules specifically designed to identify its exploitation. If an attacker can turn the flaw into a working exploit, a window opens in which the victim may still believe the system is secure even though a viable attack already exists.

Not every flaw is easy to exploit, and not every AI finding turns out to be correct. But the important change for defenders is that vulnerability discovery and validation can take place at an increasing scale. An organization that struggles to fix even known bugs must now prepare for unknown weaknesses to be discovered more quickly as well.

The use of AI in actual attacks has also been documented. On November 13, 2026, the FBI issued a warning about ARTEX, an autonomous penetration testing framework. According to the agency, various malicious actors had been using it since at least August in campaigns against organizations across different sectors.

ARTEX connects a language model to tools that can investigate systems and perform actions. The agent can analyze results and choose its next step accordingly. A human does not have to direct every action manually.

In October 2026, Microsoft described activity by a group linked to China in which ARTEX helped automate attacks against hundreds of organizations. The campaign involved the abuse of access associated with ConnectWise applications and attempts to authenticate to more than 2,250 Microsoft 365 customer environments.

In November, Microsoft described an incident in which an AI agent progressed from searching GitHub for exposed credentials to extracting data from Azure Key Vault in five minutes. This shows how quickly an exposed secret can become an actual data theft.

Anthropic has described the misuse of Claude Code in both espionage and extortion. One case disclosed in 2025 targeted at least 17 organizations. AI helped automate reconnaissance, collect credentials, penetrate networks and analyze stolen information. In a separate case, a criminal with limited programming experience used AI to help create ransomware.

Google’s threat intelligence team has also published examples of AI being used within malware. PROMPTSTEAL used a language model to generate commands for collecting data. The experimental PROMPTFLUX requested modified versions of its own code from Gemini to make detection more difficult. PROMPTLOCK was an experimental ransomware concept that used a language model to generate scripts. FRUITSHELL contained instructions intended to mislead language models analyzing malware.

These cases show several developments occurring at once. AI helps create tools, direct attacks, analyze stolen data and test ways to evade defensive measures.

That raises the question of whether existing security systems and teams can keep pace.

Attacks using AI are not automatically invisible. The published investigations themselves show that they can be detected. But detecting them requires an organization to collect the right data and connect events. An attacker may use entirely ordinary administrative tools or valid credentials. An individual action may look normal even though the sequence of actions is malicious.

My concern is whether the organization can see that whole picture. Are application activity, permission changes and cloud service usage being logged? Is anyone investigating anomalies? Does the team have the time and skills to understand what the collected data actually shows?

Improving visibility may require software development and architectural changes. Applications must record the events needed for investigations. Logs must be retained long enough and protected against tampering. Access between services must be restricted. It must be possible to contain an attack even when it exploits an unknown vulnerability.

Japan’s experience shows how serious the consequences of a cyberattack can be. Asahi confirmed that a cyberattack caused its system disruption on September 29, 2025. According to the company’s investigation, the attacker entered the network approximately ten days earlier, gained administrator privileges and explored internal systems before deploying ransomware. The company later announced both technical improvements and stronger governance arrangements.

Asahi’s published findings do not establish that AI was used in the attack or that it was connected to the current US outages. The Japanese case nevertheless demonstrates how an intrusion at a major organization can turn into a widespread operational disruption. AI can accelerate steps in such attack chains that people previously performed manually.

Taken together, these circumstances paint a troubling picture for me. Systems are aging. Known vulnerabilities remain unpatched. Experienced people are leaving. Staff numbers are being reduced in anticipation of greater efficiency from AI. Meanwhile, AI is becoming more capable of discovering and exploiting both old and previously unknown weaknesses.

The concentration of recent US outages within a short period deserves thorough investigation. In my view, a possible broader campaign of attacks using AI should be taken seriously as an investigative hypothesis. Confirming it requires technical evidence, because shared service providers, dependencies and failures with different causes can also produce simultaneous disruptions.

The same risk affects other countries, including Estonia. Our services depend on international platforms, software and suppliers. A security problem at another organization can quickly become a threat to our own operational continuity.

In my view, leaders should now be answering very specific questions. Which systems need replacing? Which known weaknesses remain unresolved? What critical knowledge left with the employees who were laid off? Has AI actually taken over their work, or is the work simply not being done? Can we detect an attack that exploits a vulnerability that has not yet been disclosed?

Treating security and experience merely as costs to be cut is becoming increasingly dangerous. As attackers work faster, defenders must also improve their ability to understand, repair and restore systems. Otherwise, we may become accustomed to major outages as a normal occurrence, with their true causes explained only after the damage has already been done.

Sources

  1. Associated Press. US technology outages and explanations from Microsoft and Chase
  2. Associated Press. The impact of Chase’s technical problem on banking services
  3. Yahoo Finance. Overview of November 2026 outages and their reported causes
  4. GAO. Critical legacy federal IT systems and their modernization plans
  5. Challenger, Gray & Christmas. Annual job cuts report for 2025
  6. Challenger, Gray & Christmas. October 2026 report, including job cuts attributed to AI
  7. ISC2. Research on cybersecurity staffing, budgets and skills
  8. ISACA. Research on cybersecurity staffing and employee training
  9. Verizon. Summary of the 2025 Data Breach Investigations Report
  10. Google Project Zero. Big Sleep’s discovery of a previously unknown SQLite vulnerability
  11. Anthropic. Mythos Preview’s vulnerability discovery and exploitation capabilities
  12. Anthropic. Project Glasswing and the discovery of longstanding software vulnerabilities
  13. OpenAI. Research into discovering, validating and patching Linux kernel vulnerabilities
  14. MARVEL. Using language models to identify vulnerabilities in hardware description code
  15. FBI. Warning about the misuse of the ARTEX framework
  16. Microsoft. ARTEX use in a large attack campaign associated with a state actor
  17. Microsoft. AI agents used to discover and abuse exposed credentials
  18. Anthropic. Investigation of an espionage campaign orchestrated with AI
  19. Anthropic. Claude misuse in data theft, extortion and ransomware development
  20. Google Threat Intelligence Group. Threat actors’ use of AI tools and malware
  21. Microsoft. ARTEX use against the financial sector and detection of related activity
  22. Asahi. Cyberattack investigation findings and measures to prevent recurrence

Home Labber who likes to build things and push it to the limits. vSphere is like Lego for adults.

“The fastest way to learn IT is 80% labbing and 20% studying theory. Just do it and have fun.” – vAndu

“If you wish to achieve worthwhile things in your personal and career life, you must become a worthwhile person in your own self-development” – Brian Tracy

VMware vExpert 2023
VMware vExpert NSX
VMware vExpert Pro
©2026 vAndu | Powered by SuperbThemes!