Using LLMs to Find and Prioritize Vulnerabilities Is No Easy Task

The integration of Large Language Models (LLMs) into the cybersecurity landscape was heralded as a panacea for the chronic talent shortage and the overwhelming volume of security alerts. However, recent implementation data suggests that using these models to identify and prioritize software vulnerabilities is proving significantly more complex than the initial hype indicated. While LLMs excel at processing vast quantities of unstructured data, they are currently struggling with the nuanced requirements of Application Security (AppSec), frequently producing high rates of false positives and failing to grasp the specific operational context necessary to distinguish a theoretical flaw from a critical exploit.

This struggle exists against a backdrop of decades of evolution in static and dynamic analysis tools. Historically, security teams relied on rules-based engines that, while predictable, were often rigid and required extensive manual tuning. The promise of the Generative AI era was a shift toward "intelligent" scanning—tools that could understand code logic and intent rather than just syntax. Industry leaders like GitHub, Snyk, and various startups have aggressively integrated LLMs to assist developers in real-time. Yet, the transition from experimental coding assistants to reliable security gatekeepers is revealing a fundamental gap between pattern recognition and true logical reasoning.

The core of the issue lies in the technical mechanics of how LLMs interpret codebases. Traditional scanners operate on deterministic logic; by contrast, LLMs are probabilistic. When tasked with vulnerability discovery, an LLM might identify a code pattern that matches a known weakness, such as an unsanitized input, without recognizing that the surrounding architecture effectively neutralizes the threat. This lack of contextual awareness means that models often flag benign code as malicious. For overstretched AppSec teams, these false positives represent a "noise tax" that can paradoxically increase their workload rather than reduce it, as human experts must still manually verify every AI-generated alert.

Furthermore, the prioritization problem remains unsolved. Effective vulnerability management requires understanding not just if a bug exists, but how reachable it is within a production environment and what the business impact of a breach would be. LLMs currently lack the visibility into a company’s broader infrastructure—such as network configurations, identity permissions, and data sensitivity—needed to make these calls. Without this metadata, an LLM treats a vulnerability in a public-facing payment gateway with the same urgency as a flaw in a siloed, offline testing tool, leading to skewed remediation priorities.

The industry implications of these limitations are significant. As organizations rush to adopt AI-driven security tools, there is a growing risk of "alert fatigue" potentially reaching a breaking point. If the market continues to prioritize the speed of AI integration over the accuracy of the output, we may see a decline in the perceived value of automated security tools. Competitively, the advantage is shifting toward vendors who can successfully combine traditional deterministic scanning with probabilistic LLM analysis—a "hybrid" approach that uses AI to explain vulnerabilities rather than just find them, thereby minimizing the burden on human analysts.

Looking ahead, the next phase of AI in AppSec will likely focus on "RAG" (Retrieval-Augmented Generation) patterns that feed specific environmental context into the model. By informing the LLM about the specific architecture and security controls of a firm, developers hope to lower false-positive rates and improve prioritization accuracy. Observers should watch for the emergence of "agentic" security workflows, where multiple models debate the validity of a vulnerability before presenting it to a human. Until these models can account for the unique context of each scan, the role of the human AppSec professional remains more critical—and more burdened—than ever.