The Cybersecurity Race Has Changed. Our Infrastructure Is Struggling to Keep Up.

AI-powered attacks, automated vulnerability discovery, software supply chains, and the growing difficulty of keeping complex systems secure and operational.
Something interesting happened on November 16.
While much of the country was focused on banking disruptions, airline outages, and the fallout from a nationwide FAA ground stop, Google published a cybersecurity report that deserves considerably more attention than it's getting.
The report is titled Agentic Threat Activity Crosses a Significant Threshold. It describes a measurable change in how cyberattacks are being conducted.
Not simply more hackers using ChatGPT to write malicious code.
Not hypothetical future attacks involving autonomous artificial intelligence.
Google Threat Intelligence Group and Mandiant documented attackers already using multi-agent AI frameworks to automate vulnerability research, coordinate offensive tools, analyze security patches, and execute portions of cyber intrusions at machine speed.
On November 13, the FBI issued its own warning about ARTEX, an autonomous penetration-testing framework being used by malicious actors.
On November 16, security researcher Scott Helme published his investigation into a compromised software service that delivered malicious JavaScript through websites belonging to 178 major organizations.
And on November 17, we learned that the National Security Agency's classified internal communications network had experienced a disruption during scheduled maintenance.
These were separate developments, with different causes and circumstances.
But taken together, they illustrate something important about the state of modern cybersecurity.
The processes of finding vulnerabilities, exploiting vulnerabilities, repairing vulnerabilities, and maintaining the systems involved are all operating under increasing technological pressure.
And those processes aren't advancing at the same speed.
The Attack Hasn't Necessarily Changed. The Execution Has.
One of the most interesting observations in Google's November 16 report is that many of the underlying methods being used by attackers are familiar.
Stolen credentials.
Unpatched software.
Misconfigured services.
Exposed administrative interfaces.
Poorly secured cloud environments.
These aren't new discoveries. Security professionals have been dealing with them for decades.
What's changing is how attackers can combine those weaknesses into an operational workflow.
Traditionally, conducting a sophisticated intrusion required substantial human involvement. An attacker might investigate a company's infrastructure, identify the software it uses, search for known vulnerabilities, test possible entry points, analyze the results, and then decide what to attempt next. Each step required time, judgment, and often specialized expertise.
Now AI frameworks can coordinate several of those activities.
Google documented the use of tools including ARTEX, CAIRN, CyberStrikeAI, and PentAGI, alongside experimentation with HexStrike AI, Agent Zero, and an inference service called ZeroHollow. These frameworks assign specialized tasks to AI agents, execute security tools, evaluate results, and adjust subsequent actions.
In some observed operations, attackers weren't simply asking an AI model how to perform a task.
They were using software that could perform parts of the task itself.
That's a significant distinction. An AI chatbot helping someone understand a vulnerability is one thing. An automated framework identifying vulnerable systems, testing attack paths, and executing available tools is something else entirely.
The specifics are worth sitting with. CAIRN implements an OODA loop where specialized agents collaborate through a shared "blackboard" and navigate autonomously from a starting parameter, like a target IP address, toward an objective. Google observed it used against targets that include the aerospace sector in the United States. CyberStrikeAI runs a Go-based backend that connects commercial and open-weight models to more than 100 security tools and YAML attack recipes. HexStrike AI wraps over 150 offensive tools and coordinates them through a dozen specialized sub-agents.
Google also documented financially motivated actors and suspected espionage operations, including a PRC-nexus group building its own automated attack framework.
One more line from that report deserves its own paragraph, because it ties directly to what Amazon documented a few days earlier. Google describes financially motivated actors profiting from LLMJacking, stealing LLM credentials and cloud resources from compromised enterprises, then reselling access to high-performance models through underground proxy services. The report calls them LLM mixers. Attackers aren't just consuming AI to run attacks. They're stealing inference and reselling it.
The report doesn't establish that every cybercriminal has suddenly acquired these capabilities. Nor does it claim these systems are infallible. Google's researchers point out that current autonomous offensive frameworks trade stealth for velocity, producing recognizable network traffic, relying on recognizable identifiers, and leaving distinct code-level traces.
But their operational use is no longer theoretical. The FBI's November 13 ARTEX advisory independently reinforces that conclusion.
The Most Important Change Might Be What Happens After a Security Patch
This is the detail in Google's report that really caught my attention.
For years, cybersecurity professionals have worried about zero-day vulnerabilities. A zero-day is a flaw attackers can exploit before a fix is available to affected users. They're dangerous because defenders have little opportunity to protect themselves.
But there's another category that deserves attention.
The n-day vulnerability.
That's a vulnerability that's already publicly known, often with a security patch available. The vendor has identified the problem. The vulnerability has been documented. A fix may already exist. The only remaining problem is getting everybody to install it.
Here's where AI creates an interesting shift.
Google documented automated frameworks capable of comparing software versions before and after patches are released. The technique is called patch diffing. The system examines what developers changed, identifies the security-relevant differences, and analyzes whether those changes reveal an exploitable weakness in the earlier version.
Attackers have used patch analysis for years. AI systems can now automate substantial portions of that research.
Google identified an exposed staging server hosting a multi-agent framework it calls exploit_agent, which integrates several frontier and local models. It compares code before and after patch announcements, analyzes the differences, drafts exploits, and executes them without a human controlling each step. GTIG describes it as the first time its researchers have observed LLM frameworks fully automating vulnerability discovery and exploitation with no human in the loop.
That's the significant development.
A security patch provides information that helps defenders protect their systems. It also provides attackers with clues about exactly what was vulnerable.
The moment a fix becomes public, a new race begins. The vendor distributes the patch. Administrators evaluate and deploy it. Attackers examine the changes and search for systems still running vulnerable versions.
And the longer those systems remain unpatched, the longer the exposure continues.
AI is reducing the amount of manual work required on the offensive side of that process. It doesn't eliminate the technical barriers to exploitation, but it changes the economics and the timing.
PHP2Shell Is a Perfect Example of the New Problem
On October 21, the NSA issued an urgent advisory about two critical PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, collectively known as PHP2Shell.
PHP is the programming language underneath enormous numbers of web applications, including WordPress.
The vulnerabilities carry a CVSS v4.0 score of 9.2 and could allow unauthenticated remote code execution on affected servers. In plain English, an attacker could execute malicious instructions on a vulnerable web server without first obtaining a legitimate account.
The NSA urged immediate patching and advised that systems unable to be adequately protected should be taken offline until remediation was possible.
Google's report reveals an additional dimension. Researchers used LLM-assisted techniques to discover the vulnerabilities. Before official patches became available, early patched code and proof-of-concept material had appeared publicly, which created opportunities to examine the differences and attempt to develop exploits. After the official patch release, hundreds of purported proof-of-concept exploits appeared online. Google reported that many were nonfunctional and likely AI-generated.
Think about what that means operationally.
Security teams weren't just dealing with critical vulnerabilities. They were dealing with a flood of exploit claims, some legitimate and some unreliable, while trying to determine whether their systems were affected and how quickly they needed to respond.
The AI acceleration wasn't confined to writing attack code. It affected the volume of material defenders had to investigate.
That's an important distinction. More security information doesn't automatically translate into faster or better security decisions. Someone still has to determine what's accurate, what's exploitable, what's relevant to the environment, and what actions can be performed safely.
Finding the Problem Is Not the Same as Fixing It
This is where the cybersecurity conversation intersects with the infrastructure outages.
There's a tendency to think of software updates as simple replacements. Something breaks. The developer releases a fix. The administrator installs it. Problem solved.
Anyone who has managed large production environments knows it isn't that straightforward.
An enterprise system may depend on hundreds of software components, configurations, integrations, and network services. Changing one component can affect another. An update that works perfectly in development may behave differently under production workloads. Security improvements can alter authentication behavior, network traffic, memory utilization, compatibility, and application performance.
When those systems support essential operations, an unsuccessful change can have significant consequences.
This isn't an argument against patching. Leaving known vulnerabilities unpatched creates substantial and well-documented security risks. It's an argument for understanding that security and operational reliability have to be managed together.
The incidents between November 12 and November 17 demonstrate why.
Microsoft: A Defect That Sat Quietly for Months
Microsoft's November 12 and 13 outage is a useful example. The disruption affected Exchange Online, Outlook, Teams, and other Microsoft 365 services.
According to Microsoft's preliminary post-incident findings, the failure combined a software defect introduced in July, increased traffic associated with service failover, and reduced available processing capacity.
The number that matters most: at the lowest point, more than 80% of the servers in the networking gateway were unresponsive.
The defect had been present in production for months without producing a widespread failure. Under the conditions encountered in November, affected gateway servers became overwhelmed.
That's an important lesson in complex systems. A defect doesn't necessarily cause an immediate failure when it's introduced. It can remain unnoticed until some later change in traffic, capacity, configuration, or operating conditions exposes it.
Microsoft's incident report attributed the disruption to technical causes, not a cyberattack or AI agents. But the engineering problem is highly relevant to cybersecurity. An update can appear successful for months before an unexpected combination of conditions exposes a latent problem.
The FAA: Modernization Meets Legacy Infrastructure
The FAA's November 13 nationwide ground stop illustrates a different version of the same challenge.
According to the agency's November 15 explanation, its legacy telecommunications infrastructure became unstable when connection traffic increased dramatically. The FAA reported approximately 1,000 connection attempts every four to five seconds.
Instability affected 17 of the agency's 22 Air Route Traffic Control Centers, disrupting communications and radar feeds sufficiently to justify a nationwide ground stop.
The initial investigation identified two locations contributing to the excessive traffic, including a communications site in El Dorado, Arkansas, where contractor L3Harris had been deploying a new router. L3Harris is running diagnostics on routers pulled from sites in Arkansas and Texas. The FAA disabled the relevant functionality on additional routers and paused further installations while investigating the root cause, and the Transportation Secretary ordered a probe while ruling out a hack.
HPE, for its part, says the Juniper router involved worked as designed. That's worth noting, because it points at the interaction rather than the component.
This wasn't a report of a successful cyber intrusion. It was a documented infrastructure problem during a modernization effort.
But notice the underlying issue. Modernizing a system requires introducing new components into an existing operational environment. Those components have to interact with infrastructure designed and deployed decades earlier, and those interactions have to work reliably under real operating conditions.
A failure at one location can become a national operational problem if the affected infrastructure supports enough critical functions.
The response also illustrates the importance of containment and recovery. The agency isolated problematic network traffic, stabilized the system, and resumed flights under established safety procedures. This wasn't merely a story about technology failing. It was also a demonstration of how essential safety systems, operational procedures, and recovery capabilities work when technology becomes unreliable.
And Then the NSA Experienced Its Own Disruption
On November 17, reports emerged that the National Security Agency's classified internal network, known as NSANet, was experiencing problems.
The disruption prompted changes to employee reporting schedules while technicians worked on the affected systems. The NSA subsequently issued an official statement explaining that the agency had experienced temporary degradation during scheduled routine maintenance. It said some internal systems were affected, but mission-critical capabilities remained operational.
The agency also emphasized its redundant infrastructure and ability to maintain essential operations.
That's a useful distinction. Not every system failure means an organization has lost its ability to perform its mission. A sophisticated cybersecurity organization isn't immune to the ordinary operational difficulties involved in maintaining complex technology.
The NSA incident wasn't publicly established as a cyberattack. Nor is there evidence that AI-assisted patching caused it. The documented explanation is a disruption during routine maintenance.
That's enough to make it relevant. Even organizations built around cybersecurity expertise must manage the same fundamental engineering challenge: changing important systems without unnecessarily disrupting the services those systems provide.
There's Another Problem: A Failure Doesn't Stay Where It Starts
The November 16 banking disruption provides a particularly clear example.
JPMorgan Chase identified a software bug affecting its internal systems and delaying transactions, including some Zelle payments. The bank stated it wasn't a cyberattack.
But customers interacting with other financial institutions also reported problems. Some transactions involving Chase were delayed even when the other institution's systems were functioning normally. Chase's own position was careful: where it was party to a transaction involving other banks, those banks may have been impacted, but this wasn't a systemic issue.
The customer sees a failed payment or an unavailable transaction. The underlying cause may exist somewhere else entirely.
It doesn't require multiple independent system failures to produce problems across multiple organizations. One significant dependency can be sufficient.
This is where reliability engineering and cybersecurity start sharing the same fundamental concern.
How far can a problem spread beyond the system where it originated?
Engineers call this the blast radius. In traditional systems, the objective is to isolate failures so a problem in one component doesn't disable everything connected to it. Cybersecurity applies a similar principle. If one account is compromised, that account shouldn't automatically provide access to an entire organization. If one application is vulnerable, the attacker shouldn't gain unrestricted access to everything else.
Which brings us to Optimizely.
The Optimizely Compromise Shows What Happens When Trust Is Shared
On November 16, Scott Helme of Report URI published the results of his investigation into malicious JavaScript delivered through the Optimizely Web experimentation platform.
He fetched 1,611 distinct Optimizely snippets and checked each one. He found 188 infected snippets belonging to 178 organizations. The list includes Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Air France-KLM, Ledger, and Peacock.
The malicious code was multi-purpose: it swapped cryptocurrency wallet addresses for the attacker's, skimmed payment cards, stole cookies and screenshots, and on iPhones and iPads delivered two iOS exploit kits called Coruna and DarkSword. Verified cryptocurrency theft landed in the attackers' wallets, roughly $7,040 within about nine hours.
The payload was rebuilt 26 times during the day. Features were added and wallets were added, but nothing was removed.
But the most important architectural detail is how the malware reached so many websites.
The attacker didn't need to compromise 178 organizations. They wrote code into the Project JavaScript field of customer projects inside Optimizely. The rest of each snippet was untouched. The file was still served from cdn.optimizely.com, with Optimizely's certificate, from a host that every affected site already trusted. This wasn't cache poisoning; the CDN served exactly what it was given.
Optimizely's own snippets were infected too.
So the legitimate code was already trusted by websites using the service. When malicious functionality was introduced into that delivery mechanism, the affected sites passed it to visitors through their normal page-loading processes.
During a subsequent television interview, Helme explained that the malicious software wasn't designed to replicate itself across networks. The distribution came from the shared supplier relationship.
That's an important difference. The attack didn't become widespread because the malware was independently infecting hundreds of corporate systems. It became widespread because an existing software service already had the necessary distribution.
Attackers don't always need more sophisticated exploits to achieve greater reach. Sometimes the architecture of the target environment provides the amplification.
Three Different Problems, One Increasingly Important Discipline
Looking across these incidents, I see three distinct engineering challenges.
The first is offensive acceleration. Google and the FBI documented malicious actors using AI-assisted frameworks to automate activities that previously required substantial manual effort. That affects how quickly defenders must discover, assess, and respond.
The second is operational complexity. The Microsoft, FAA, and NSA incidents illustrate the difficulties of maintaining and modifying sophisticated systems while keeping essential operations available. That challenge exists independently of malicious activity.
The third is dependency concentration. The Chase and Optimizely incidents demonstrate how an issue originating in one system or supplier can affect numerous other organizations. One involved a software failure. The other involved a malicious compromise. Different causes, but both show why the relationships between systems matter.
These problems shouldn't be collapsed into one category. A software failure isn't automatically a cyberattack. An AI-powered exploit framework doesn't explain an airline outage. And the fact that multiple failures happen within a short period doesn't establish a coordinated operation.
But they aren't unrelated engineering concepts. They intersect in the systems that security professionals and infrastructure operators are responsible for protecting.
We're Going to Have to Rethink What Good Cybersecurity Looks Like
Traditionally, cybersecurity performance has been measured through familiar questions. How many vulnerabilities were discovered? How many patches were installed? How quickly were alerts resolved? How many attacks were blocked?
Those measurements remain useful. But they don't tell the whole story.
A company can install patches rapidly and still introduce a serious production failure. It can deploy sophisticated security software and still depend on a vulnerable outside supplier. It can prevent an intrusion and still suffer a major outage because an internal component failed. It can have excellent backups and remain exposed to an attacker using valid stolen credentials.
The problems overlap, but the controls required to address them aren't identical.
Modern cybersecurity has to account for both preventing compromise and maintaining operations. That means understanding not only whether a vulnerability exists, but how it can be exploited, what systems depend on the affected component, how a fix can be introduced safely, and what happens if the fix or the component fails.
It also means being able to identify and contain a problem before it spreads through every connected service.
For administrators and engineers, this puts renewed importance on staged deployments, configuration management, dependency inventories, segmentation, monitoring, recovery testing, and rollback procedures.
Not particularly glamorous technologies. But absolutely essential ones.
AI has legitimate applications on the defensive side as well. Automated code review, vulnerability detection, incident triage, and security testing can help defenders identify problems sooner and prioritize their response.
The goal isn't to avoid automation. It's to build automation that operates within reliable controls. Faster decisions and faster changes are only useful when the results can be trusted.
The Real Shift Is in the Time Available to Respond
I think this is the central finding connecting Google's research with the larger cybersecurity environment.
The underlying vulnerabilities haven't all become new. Software still contains defects. People still make mistakes. Credentials still get stolen. Dependencies still fail.
The change is in the speed and scale at which some of those weaknesses can be discovered and exploited.
At the same time, repairing production infrastructure remains a complicated process. Organizations have to evaluate patches, test compatibility, coordinate changes, maintain continuity, recover from unsuccessful deployments, and manage systems that depend on outside providers whose operations they don't directly control.
Some of those activities can be automated. Others still require careful engineering decisions and operational safeguards.
The important distinction is that the time required to exploit a vulnerability and the time required to safely remediate it aren't necessarily shrinking together.
Google's findings demonstrate meaningful advances in automated offensive capability. The infrastructure incidents of November 12 through 17 demonstrate the continuing importance of reliability engineering, operational controls, and resilient system architecture.
Neither finding proves the other caused anything. They don't need to.
Together, they show why cybersecurity can no longer be evaluated solely by how effectively an organization prevents intrusions. It must also be evaluated by how well that organization understands its dependencies, manages changes, limits failures, and maintains essential operations.
We've spent decades optimizing technology for connectivity, availability, convenience, and efficiency. Now we're operating that technology in an environment where vulnerability research and parts of the exploitation process can increasingly run at machine speed.
That's a significant change in the operating conditions. And the response requires more than finding vulnerabilities faster. It requires systems designed to tolerate failures, contain compromises, accept necessary changes safely, and recover when something inevitably goes wrong.
The next major challenge in cybersecurity isn't just discovering what's vulnerable. It's securing increasingly interconnected infrastructure without losing the ability to operate it reliably.
That's where the research, the incidents, and the engineering lessons of those six days come together.
Sources
Google Threat Intelligence Group, Agentic Threat Activity Crosses a Significant Threshold, November 16, 2026. https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
FBI, FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework, November 13, 2026. https://www.ic3.gov/CSA/2026/261113.pdf
NSA, Critical Patches Recommended for PHP Vulnerabilities, October 21, 2026. https://www.nsa.gov/Press-Room/Press-Releases-Statements/Press-Release-View/Article/4630238/nsa-advisory-critical-patches-recommended-for-php-vulnerabilities/
Scott Helme, World's Largest Brands Infected With Malware Via Optimizely, November 16, 2026. https://blog.report-uri.com/worlds-largest-brands-infected-with-malware-via-optimizely/
Microsoft 365 preliminary post-incident findings, reported by BleepingComputer. https://www.bleepingcomputer.com/news/microsoft/microsoft-says-months-old-bug-caused-global-microsoft-365-outage/
FAA, Nationwide Ground Stop: Questions and Answers, November 15, 2026. https://www.faa.gov/newsroom/nationwide-ground-stop-questions-and-answers
NSA, Official Statement on Network Issues, November 17, 2026. https://www.nsa.gov/Press-Room/Press-Releases-Statements/Press-Release-View/Article/4650245/nsa-statement-on-network-issues/
Washington Post, Banking disruptions involving JPMorgan Chase, November 16, 2026. https://www.washingtonpost.com/business/2026/11/16/bank-outage-chase-wells-fargo-truist-zelle/ea7727d6-e215-11f1-a99e-b3e3805c6b8b_story.html
Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.





Comments