# AI in Cybersecurity: Making Vulnerability Backlogs More Actionable

- **Canonical URL:** https://www.probackend.com/ai-open-source-vulnerability-patching/ai-in-cybersecurity-making-vulnerability-backlogs-more
- **Section:** Vulnerabilities & Exploits
- **Published:** 2026-11-16
- **Summary:** Learn how AI in cybersecurity can help prioritize reachable open-source vulnerabilities, and what the reported 3% attackability figure does—and doesn’t—show.

---

## Artificial intelligence in cybersecurity and the vulnerability queue

A vulnerability alert is not the same thing as an exploitable weakness in your application. That distinction matters when teams are staring at thousands of dependency findings and a fresh critical CVE has just landed. A 2022 ShiftLeft report, covered by [Dark Reading](https://www.darkreading.com/application-security/open-source-software-bugs--attackability), argued that only a small share of vulnerable open-source components were actually reachable in the applications examined. Its headline figure was about 3% attackable.

That is a useful prompt to investigate context—not permission to ignore the other 97%. The result is vendor-reported research, not a universal measurement of every software portfolio, and “not attackable” under an analysis does not mean “harmless forever.” Still, if every dependency alert gets the same urgency, teams spend scarce engineering time proving negatives instead of fixing the paths that matter.

## What is AI in cybersecurity?

Artificial intelligence in cybersecurity describes computational methods that help analyze security-relevant data, detect patterns, and support decisions. In application security, tools may analyze dependency inventories, code paths, and data flows to help determine whether a vulnerable component is loaded and reachable. That kind of analysis can make a huge queue more intelligible.

But the 3% finding is not evidence that AI itself discovered or prevented those flaws. It is about attackability analysis and prioritization. The practical lesson for AI cybersecurity is to keep the claim modest: automation can help teams gather context and rank work; people still need to validate assumptions, weigh business impact, and decide what to patch or mitigate.

## Reachability is more than package presence

A scanner can identify a known vulnerability in a library included in a software bill of materials. That is an important starting point, but it leaves questions unanswered. Is the dependency loaded? Does the application call the vulnerable function? Can attacker-controlled input reach it through the relevant data flow? Is the package actually used in the deployed application?

The Dark Reading account describes attackability assessment as considering whether the affected package is loaded and in use, whether it lies on an attacker-controlled path, and whether it can be reached through data flows. This resembles a focused form of threat modeling: not merely “does this component exist?” but “is there a plausible path from an attacker’s input to the vulnerable behavior?”

Severity remains relevant. Reachability adds context; it does not cancel the potential impact of a flaw. A critical issue with a confirmed route into vulnerable code deserves attention. A vulnerable package that appears unreachable may be lower priority today, but changes to application behavior, configuration, or dependencies can alter that assessment.

## How AI is used in cybersecurity triage

In vulnerability management, automated analysis can help correlate package findings with application behavior. It can reduce the manual effort of checking whether a dependency is used, and help surface routes through code and data flows that merit review. That is one answer to how AI is used in cybersecurity: to sift and connect evidence at a scale that would be tedious for a person to inspect one alert at a time.

The tool’s output should be treated as a decision aid, not an oracle. Models and analysis engines depend on the code, configuration, and runtime context available to them. An incomplete inventory or a missed execution path can produce a misleading confidence signal. Teams should preserve the underlying evidence and make it possible to revisit a finding when the application changes.

The report said developers saw a 97% reduction in false-positive library-upgrade tickets when they considered attackability for packages with critically rated vulnerabilities. It also reported that 76% of vulnerabilities were fixed within 12 days, described as two sprints, and a 37% reduction in mean time to remediate. These are reported outcomes from the study, not guaranteed effects for every organization or proof that a single analysis technique caused every improvement.

## Why the 3% figure needs careful handling

A percentage can sound more definitive than the underlying question. The reported finding concerns the vulnerabilities and applications examined in the study; it does not establish that 97% of all open-source flaws worldwide are benign, nor that organizations can safely leave them unattended. The value lies in reducing noise and directing investigation—not turning a study statistic into a blanket exception policy.

The Log4j experience is a useful caution. Dark Reading’s reporting notes that 96% of vulnerable Log4j dependencies in the report were judged not attackable, while also describing industry concern that paths to exploitation may be difficult to anticipate. Mark Curphey, an application-security advocate quoted in the article, cautioned that determining which dimensions make something attackable can be complicated in modern applications. A path that looks obscure can still matter.

That tension is healthy. Security teams need prioritization, and they need humility about what their analysis can see. A “not reachable” result should have a scope, a date, and evidence behind it. It should not become a permanent waiver.

## Dependencies make prioritization unavoidable

Modern applications rely on open-source components directly and through other dependencies. As the Dark Reading report notes, teams may encounter new CVEs multiple times a month for applications with many dependencies. Multiply that across an organization’s portfolio and the alert queue grows quickly.

Yet fixing a library is not always a simple version bump. The reporting quotes developers pointing out that upgrades can require broad testing for security, functionality, and quality; code may need refactoring, and a new version may not be backward-compatible. That is why ranking work matters. Teams must choose where to spend validation and release effort without pretending that engineering constraints eliminate risk.

A sensible workflow starts with inventory and severity, then adds reachability evidence and application context. Prioritize confirmed attacker-accessible paths, assess exposure and impact, and plan a patch or compensating control. For findings assessed as unreachable, record the reasoning, monitor for changes, and revisit when code or configuration shifts. Keep the ability to escalate quickly when new evidence appears.

## Use automation to sharpen judgment

The strongest case for AI in cybersecurity is not that it makes judgment unnecessary. It is that good analysis can spare teams from treating every alert as equally urgent. Reachability can turn a dependency list into a more useful map of risk, provided the map is not mistaken for the territory.

So use the 3% finding as a challenge to improve triage, not as a target to chase. Ask what the analysis actually observed, which paths it covered, and what would change the conclusion. Patch what is demonstrably exposed; document and review what is not. And when an unfamiliar flaw becomes the next Log4j-scale fire drill, retain enough caution to question yesterday’s assumptions.

![artificial intelligence in cybersecurity and the vulnerability queue](https://assets.probackend.com/136ba86a-51d2-4d5d-a0d6-714d654367c3?token=ast_7b9ee8975a4b187969d396286e3f9717c42f2331)
