COMMENTARY: Security teams have spent years prioritizing vulnerabilities based on CVSS severity or the potential of attackers to exploit them in isolation.
But they have also struggled to understand the complete picture of how attackers can combine multiple weaknesses together across assets to reach crown jewels.
[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]
This lack of big-picture visibility will get a lot worse. Human hackers have limited time and resources, and every additional step introduces another opportunity to fail or trigger an alert. Attackers generally pursue the easiest available route to their target because spending hours on an uncertain exploit rarely makes economic sense.
AI agents change that calculation because they can spend hours exploring paths that a human attacker would abandon after a few minutes. They can test multiple routes simultaneously, recover from failed attempts and work through a chain of seemingly insignificant weaknesses until they reach something valuable.
Security teams now have to contend with a growing number of attack paths that their existing prioritization methods may overlook.
Multiple attack paths for every one a human would pursue
At Cogent Security, we recently analyzed millions of enterprise assets and billions of security finding across dozens of large enterprises. In August 2026 alone, the average enterprise accumulated dozens of new attack paths to its most sensitive systems. Only 25% of these paths were within the practical capabilities of human attackers. The remaining 75% exceeded those limits, but were still viable for autonomous AI agents.
The rate at which these machine-viable paths appeared has quadrupled year over year, compared with a more modest increase for human-viable paths. The number of these paths exploitable by AI has grown in part because of the advancement of capabilities in AI models, first seen in Mythos and now in open weight models like GLM 5.3.
These trends describe potential attack paths rather than actual intrusions, but they show how much exposure security teams can overlook when they evaluate vulnerabilities according to the capabilities and behavior of human attackers.
The paths themselves also run considerably deeper. The median machine-viable path crossed eight assets and five trust boundaries, compared with four assets and two trust boundaries for human-viable paths.
Consider an attacker who compromises an employee's laptop. Reaching a sensitive production database might require obtaining a service account credential, exploiting an overly-permissive identity role, crossing into a cloud environment and finding another configuration weakness that permits access to the database.
Finding this path requires trial and error. A human attacker might give up halfway through discovering it, especially if several steps have uncertain outcomes. An autonomous agent can continue testing the chain while other agents investigate alternatives, making vulnerabilities that once seemed too difficult to exploit considerably more attractive.
Why low-severity vulnerabilities deserve another look
Most enterprise security programs still rely heavily on CVSS severity scores to determine what gets fixed first. Given the volume of vulnerabilities in a large organization, security teams need a way to distinguish urgent problems from those that can wait.
The trouble comes when teams evaluate findings individually without considering how an attacker could connect them.
Our research found that 7% of critical attack paths consisted entirely of low- and medium-severity vulnerabilities. Each weakness might fall below an organization's remediation threshold, yet an attacker who successfully chains them together could reach a critical asset.
These paths also tend to remain available for a long time. Eighty-three percent of machine-viable paths persisted for more than 30 days, compared with half of human-viable paths.
That gives attackers a substantial window to discover and exploit them. Organizations can spend weeks addressing high-severity vulnerabilities while leaving intact a route to sensitive systems built entirely from findings that their security programs consider relatively unimportant.
Focus on the paths, not just the individual risks
Security teams need to understand how risks connect across their environments, which requires bringing together information from vulnerability scanners, endpoint tools, identity platforms, and cloud security products.
In our research, reconstructing the median critical attack path required correlating information from five different security tools. Each tool could identify part of the problem, but none could independently reconstruct the complete route.
Security teams should use that combined information to determine which weaknesses give attackers access to their most valuable systems and where they can interrupt those routes with the least effort.
Sometimes patching a critical vulnerability will accomplish that. In other cases, removing an unnecessary permission or correcting a configuration will eliminate several potential attack paths at once, even when the individual finding carries a relatively low severity score.
Organizations will always have more vulnerabilities than they can immediately fix, so they should make it a priority to identify which changes eliminate the greatest amount of exploitable risk.
As autonomous agents make longer and more complicated attacks practical, security teams need to reconsider how they decide which vulnerabilities deserve attention. Otherwise, they risk spending their limited resources fixing individual weaknesses while leaving attackers a viable route to the systems they most need to protect.
Vineet Edupuganti, co-founder and CEO, Cogent Security
SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
