The Cybersecurity Crisis That Has Nothing to Do With AI

A hospital system in the Midwest discovers that someone has been inside its network for 347 days. The intruder did not use a zero-day exploit. It did not deploy an AI-generated deepfake voice to trick the help desk. It entered through a VPN appliance that was three months past its patch cycle, using credentials that had not been rotated since the previous administration. By the time anyone noticed, patient records for 1.2 million people were being negotiated on forums that require an invitation to even browse.

That incident, part of a pattern documented in the CISA's updated RESURGE malware analysis, published in February 2026, is worth holding onto as the cybersecurity industry enters its annual cycle of threat report season. The reports arriving this quarter paint a vivid picture of an accelerating arms race. The CrowdStrike 2026 Global Threat Report, released in late February, documented an 89% year-over-year surge in AI-enabled adversary operations. The average breakout time, the window between initial compromise and lateral movement, collapsed to 29 minutes. The IBM 2026 X-Force Threat Intelligence Index, published the following day, reported a 44% increase in attacks targeting public-facing applications, many enabled by weak or missing authentication controls. The headlines write themselves: AI is eating security, and the defenders are losing.

But the same reports contain a quieter finding that cuts against the dominant narrative. CrowdStrike's own analysis draws a careful distinction: AI is accelerating established tactics rather than creating entirely new ones. IBM's data shows that the single largest category of successful intrusions still involves exploitation of known vulnerabilities for which patches already exist. The threat is real, but the story being sold, that a new class of hyper-intelligent AI attackers has rendered traditional defenses obsolete, does not match what the data actually shows. What is happening is both more prosaic and, in some ways, more damning for organizations that have spent years deferring basic maintenance.

What AI Actually Changes About an Attack

To understand why this distinction matters, it helps to walk through the mechanics of a modern intrusion. Before generative AI, a ransomware operator looking to breach a mid-sized logistics company had to write phishing emails manually, research the organization's structure from scratch, and move laterally by guessing passwords or hoping an administrator had left a remote desktop port exposed. Each step was slow, labor-intensive, and left forensic traces. The operator might compromise one or two organizations per campaign cycle.

Generative AI changes the throughput, not the path. The same operator can now use a language model to draft phishing copy in the voice of the CEO, customized to the latest internal memo the attacker scraped from a compromised SharePoint site. AI tools can scan codebases for logic flaws faster than a human reverse engineer. They can generate variations of credential-stuffing attempts that evade pattern-based detection. The IBM report notes that attackers are leaning heavily on AI to identify and validate vulnerabilities faster. The CrowdStrike report finds that even less sophisticated threat actors can now execute攻击 that previously required specialized expertise.

The result is an acceleration of the same kill chain that has existed for two decades: reconnaissance, initial access, privilege escalation, lateral movement, data exfiltration, impact. The steps did not change. The speed did. A process that once took a week now takes an afternoon.

The Problem That Did Not Change

Here is where the data from both reports converges on an uncomfortable finding. The IBM X-Force team analyzed every incident it responded to in 2025 and found that the root cause in a majority of cases was not an AI-generated exploit that bypassed all known defenses. It was a vulnerability listed in the Common Vulnerabilities and Exposures database, often months or years old, that had not been patched. It was an application exposed to the internet with default credentials still active. It was a cloud storage bucket configured to allow public write access.

CrowdStrike's report tells a similar story through a different metric. The company tracked the most common initial access vectors across all incidents in 2025. The top entries were, in order, exploitation of public-facing applications, compromised credentials, and phishing. These are the same three vectors that have topped the list for the better part of a decade. AI did not invent them. AI made each one faster and harder to detect, but the defender's fundamental task, patch the application, rotate the credential, filter the phish, remained structurally identical.

The RESURGE malware that CISA updated its analysis on in February is a case study in why this distinction matters. RESURGE is a sophisticated implant that exploits a stack-based buffer overflow vulnerability in Ivanti Connect Secure appliances, tracked as CVE-2025-0282. It is stealthy, persistent, and can remain dormant for extended periods. But its entry mechanism is a known vulnerability in a specific class of devices. Organizations that prioritized patching their Ivanti gateways when the advisory dropped were protected. Those that did not, or that lacked an inventory system capable of telling them whether they owned an affected appliance, remained exposed.

The Infrastructure That Leaks

The problem is not that defenses are failing against AI. The problem is that the defenses themselves are often not deployed. A 2025 survey by an industry consortium found that roughly 40% of mid-market companies do not maintain an accurate inventory of their internet-facing assets. You cannot patch what you do not know exists. You cannot rotate credentials for a service that no one in the organization remembers deploying. Attackers know this. They scan the same IP ranges, using the same tools available to any security researcher, and they find the forgotten VPN, the expired certificate, the SSH port left open on a server that was supposed to be decommissioned two years ago.

This is where AI does become a genuine force multiplier, but not in the way the headlines suggest. Attackers are not using AI to invent novel cryptographic attacks or to brute-force AES-256. They are using AI to automate the reconnaissance phase so thoroughly that they can identify an organization's full attack surface, including the forgotten assets, in hours instead of weeks. The CrowdStrike report documents cases where attackers used AI tools to scan public code repositories for embedded API keys and cloud credentials, then used those credentials to access internal systems within minutes of discovery. The breach method is not new. The speed at which it can be executed at scale is.

And when attackers find a way in through a known vulnerability, they are increasingly using AI to accelerate the lateral movement phase. Instead of manually enumerating Active Directory trusts or testing service accounts one by one, they feed the initial foothold into an AI agent that maps the network topology, identifies privileged accounts, and suggests the shortest path to the domain controller. Palo Alto Networks' Unit 42 team, in its 2026 Global Incident Response Report, found that the median time from initial compromise to domain-level access dropped by over 60% in incidents involving AI-assisted tools.

What Protection Actually Requires Now

If the threat is acceleration of existing attack patterns rather than emergence of fundamentally new ones, then the defensive playbook shifts in a particular way. It does not require abandoning existing security programs in favor of AI-powered defense platforms. It requires executing the basics with a rigor that most organizations have never achieved, and then layering selective automation on top of that foundation.

Patch management is the obvious starting point. The IBM report documents that a significant percentage of successful intrusions involved exploitation of vulnerabilities for which patches had been available for more than six months. The median organization takes over 100 days to patch critical vulnerabilities in internet-facing systems. In an environment where the average breakout time is 29 minutes, a 100-day patch window is not a speed problem. It is a structural failure of prioritization.

Credential hygiene is the second lever. The CrowdStrike report notes that the majority of intrusions involving compromised credentials could have been prevented by enforcing multi-factor authentication on all externally accessible services. Not sophisticated MFA that requires behavioral biometrics or hardware tokens. Standard time-based one-time passwords. The kind of MFA that has been widely available for over a decade. Organizations that have not deployed it everywhere are not being outsmarted by AI. They are being outwaited by attackers who know exactly which services lack the extra step.

Network segmentation is the third. In incidents where the attacker reached the domain controller within minutes, the common thread was flat network architectures. The backup server, the HR file share, the domain controller, and the public-facing web application were all on the same network segment with no firewall rules between them. Once the attacker got through the front door, there were no internal doors. Segmentation does not require a PhD in network engineering. It requires a diagram of traffic flows and the discipline to enforce rules that limit lateral movement.

The Distraction of the New

The risk of the current narrative around AI-driven cyber threats is that it tempts organizations to chase shiny solutions while neglecting the infrastructure work that actually determines breach outcomes. A board that hears that AI-powered attackers are now impossible to stop without AI-powered defense platforms may approve a budget for an expensive AI-driven security tool while simultaneously cutting the budget for the patching team. The tool will generate impressive dashboards showing threats detected and blocked. But if the attacker enters through a VPN appliance that was never patched, the dashboard is irrelevant.

There is a parallel here to the way supply chain disruptions were misdiagnosed during the peak of the pandemic. Many companies rushed to implement AI-powered forecasting tools while the actual bottleneck, container ship scheduling and port congestion, remained unaddressed. The shipping bottleneck we documented was an infrastructure problem masquerading as a data problem. The cybersecurity situation in 2026 is similar: the perimeter is not failing because attackers have discovered magic bullets. It is failing because too many organizations have not locked the doors that have been unlocked for years.

None of this is to minimize the genuine risk posed by AI-enabled adversaries. The 89% surge documented by CrowdStrike is real. The 29-minute breakout time is alarming. The democratization of advanced attack techniques means that organizations once considered too small or too obscure to be targeted now face the same threat landscape as Fortune 500 companies. But the correct response to acceleration is not to abandon the fundamentals in favor of experimental defenses. It is to run the fundamentals at machine speed, automated where possible, enforced by policy where not.

The RESURGE Lesson That Keeps Repeating

CISA's updated analysis of the RESURGE malware, released in February 2026, contains a section that few press summaries highlighted. The malware, despite its sophistication, was detected in most cases not by advanced AI-powered endpoint tools but by organizations that had implemented the detection signatures CISA had published the previous year. The organizations that found it were the ones that had operationalized the indicators of compromise, the specific file hashes, IP addresses, and registry keys, into their monitoring systems. The organizations that did not find it were the ones that had not.

The question that the market is not asking loudly enough is whether the cybersecurity industry has a fundamental incentive problem. The vendors who sell AI-powered detection platforms benefit when the narrative emphasizes the novelty and sophistication of the threat. A story about an unpatched VPN appliance does not sell an eight-figure enterprise license renewal. A story about an AI-generated polymorphic malware that evades all signature-based defenses does. The result is a collective overinvestment in detection technology at the expense of prevention fundamentals.

CrowdStrike and IBM, to their credit, have been transparent about where the actual risk lies. The CrowdStrike report explicitly states that AI is accelerating existing tactics rather than creating new attack categories. The IBM report titles its core finding plainly: basic security gaps are leaving enterprises exposed. The data is public. The question is whether organizations will read beyond the headlines.