JFrog: AI Coding Tools Are Generating Vulnerabilities 10 Times Faster Than They're Being Fixed

The vulnerability management playbook for most security teams is built around severity filtering: triage by CVSS score, prioritize critical and high, defer medium, ignore low. That model assumes a stable inflow volume that a team with a defined capacity can work through over time. The 2026 data from JFrog’s security report makes the case that the model has broken.

The specific pressure point: AI coding tools. JFrog’s data attributes approximately 35 new CVEs per month directly to AI-generated code. Against a background where total disclosed CVEs grew 92% between 2024 and 2025, the AI-generated component isn’t a footnote — it’s a meaningful contributor to a volume problem that severity-based filtering can no longer contain.

The Core Finding: Bigger Backlogs, Different Shape

The 92% year-over-year increase in disclosed CVEs represents a structural change in the threat environment, not a measurement artifact. The number of CVEs that were disclosed and needed some organizational response roughly doubled. Most security teams did not double their capacity to respond.

CISA’s formal recommendation — published alongside this data — is to abandon severity-based filtering as the primary triage mechanism. The argument: when volume overwhelms capacity, filtering by severity doesn’t reduce the backlog, it selects which portion of the backlog gets oldest fastest. High-severity vulnerabilities get addressed; medium and low accumulate; exploitable medium-severity vulnerabilities sit untouched because the priority scheme excluded them.

The Attacker Timing Problem

Attackers scan for new CVEs within 15 minutes of public disclosure. This is the organizational mismatch that makes volume management a genuine security posture problem, not just a process efficiency problem.

A security team prioritizing by severity and working through critical/high vulnerabilities first can take weeks to reach a medium-severity CVE that is actively being exploited. The 15-minute attacker response window creates a consistent gap between disclosure and remediation that severity-based filtering doesn’t close — it just concentrates the gap at the lower end of the priority queue.

Where AI Code Fits

The 35 CVEs/month from AI-generated code isn’t a ceiling — it’s the 2026 baseline as AI coding adoption accelerates. JFrog’s analysis attributes this to a consistent pattern in how large language models generate code: they reproduce patterns from training data that includes vulnerable implementations, particularly around:

• Insecure cryptography defaults (deprecated algorithms, predictable key generation)
• Input validation gaps that model from existing examples with the same gaps
• Dependency selection that pulls in older, vulnerable versions that appear frequently in training data
• Authentication and authorization patterns that are technically functional but miss edge cases

The security problem with AI-generated code isn’t that the code is obviously wrong — it’s that the code passes review because it looks like every other implementation and the vulnerabilities are subtle.

The CISA Alternative to Severity Filtering

If severity-based filtering is insufficient, what replaces it? CISA’s position is exploitability-informed prioritization: move vulnerable-in-wild evidence, known exploit kits, and reachability analysis (does the vulnerable code path actually execute in your deployment?) to the top of the triage signal over CVSS score alone.

This requires different tooling than severity-based filtering. CISA’s Known Exploited Vulnerabilities (KEV) catalog is the starting point — a maintained list of CVEs with confirmed active exploitation. JFrog’s platform integrates the KEV catalog as a prioritization signal alongside static CVSS scoring.

Reachability analysis is the harder piece: does your codebase actually call the vulnerable function, with inputs that reach the vulnerable path? A high-CVSS vulnerability in a library function that your code never calls is a lower priority than a medium-CVSS vulnerability in a code path that executes on every user request. Static analysis tools that produce reachability graphs let teams filter the 92%-larger backlog more precisely than CVSS score alone.

The Platform Engineering Implication

For platform and security engineering teams: the 10x vulnerability generation rate from AI tools is an argument for integrating security scanning earlier in the AI-assisted development loop — at the code generation stage, not just at the PR review stage. Tools that scan AI-generated code before it enters the review queue can interrupt the 35-CVE/month cycle at the source rather than processing the output.

JFrog’s platform supports pre-merge scanning that applies this model. Whether you use JFrog or an alternative, the architecture principle is the same: move the security gate earlier in the pipeline for AI-generated code specifically, because the traditional assumption — that a human wrote the code and a human reviewing it would notice obvious issues — doesn’t apply in the same way.

Source: JFrog Blog

Content created with AI assistance and reviewed for accuracy.