JFrog CISO: AI Is Reshaping the Vulnerability Backlog, Not Just Growing It

A backlog that grows faster than release cycles

The post opens with figures from outside sources to describe the pressure on security teams. The Cloud Security Alliance reports that AI-assisted developers generate code commits three to four times faster than their peers but introduce security findings at ten times the rate. That, the post says, creates security debt that outpaces remediation.

Georgia Tech's Vibe Security Radar tracked 35 CVEs in one month alone that were directly linked to AI coding tools. Researchers estimate the true number in open source packages is five to ten times higher. The UK's NCSC warned of a "wave of patches" this spring and attributed it to years of prioritizing rapid delivery over vulnerability debt. According to the post, disclosed vulnerabilities surged 92% in 2025, even before the latest acceleration.

Severity-based triage is no longer enough

The author calls the qualitative change the more significant one. For two decades, teams patched criticals, scheduled highs and let mediums and lows wait. The post calls this approach obsolete.

The Australian Cyber Security Center's April guidance on frontier models advised organizations to reassess risk tolerance and patch vulnerabilities regardless of severity. The concern is chained vulnerabilities, where lower-rated flaws combine into a more serious threat. The post says new AI agents excel at this. Its example: a medium-severity information disclosure and a low-severity path traversal, both previously deprioritized, can now be exploited together.

The response window has also narrowed. Unit 42 reports attackers scanning for new CVEs within fifteen minutes of publication. A critical pre-auth RCE in the Marimo notebook project was exploited within ten hours of disclosure, using only the advisory text.

For software producers, the post says, medium and low issues ship to customers and get flagged by their scanners. They may trigger support tickets, security alerts or contractual remediation obligations. Customers could filter by severity, but zero-tolerance policies mean they rarely do. Engineering teams therefore have to fix the full spectrum of vulnerabilities with no increase in resources.

The CISO's position and the SolarWinds lesson

Proofpoint's 2026 Voice of the CISO found that 79% of CISOs are expected to manage AI risk without additional resources, according to the post. The author then recounts a conversation at this year's swampUP conference with Tim Brown, who was SolarWinds' CISO during the SUNBURST attack.

That attack was caused by 3,000 lines of malicious code inserted into the build environment, bypassing source control. The response included a pipeline overhaul with three separate builds for development, staging and production. Access was reduced from 400 to 36 to 2 people, binary comparisons across builds were added to detect tampering, and new features were paused for six months. The SEC considered filing charges against Brown as an individual. The post says the case was dropped after 3 years, and that the stress led to a heart attack.

The author draws a lesson about misplaced trust: assuming code is safe or systems are properly configured without verification. Controls that once ran on a six-month audit cycle must now run continuously. In the agentic era, threats include autonomous agents as well as rogue employees, which makes non-human identity management and agent sandboxing critical.

What the post recommends, and JFrog's pitch

The central capability, according to the post, is precise and continuous knowledge of software artifacts, their deployment and ownership, plus a response plan if they are compromised. Without it, the author says, meeting regulatory deadlines would be impossible. These include the 24-hour early warning under the EU Cyber Resilience Act and the 72-hour notification to ENISA for actively exploited vulnerabilities in products sold in the EU.

The author says only the CISO is positioned across executive, engineering and customer domains to advocate for a unified plan, but that responsibility is shared. Brown stressed communication and collaboration.

On the product side, JFrog says it is not building another scanner that reports issues after the fact. It describes a platform where artifacts, provenance, governance, dependencies, policy and release paths sit in one centralized system of record. As agents write, review and release code, the post says, the platform must be the control point. That means managing non-human entities, ensuring verifiable provenance, enforcing policy at pipeline speed and maintaining an immutable build record. The company claims this approach lets teams ship faster while keeping standards they can present to regulators, customers and the board.