Between January and July 2026, the National Vulnerability Database (NVD) cataloged 45,207 flaws, a volume putting the year on pace to double the total count of 2025. This surge is not merely a byproduct of better reporting; it is the primary signal of a widening structural gap where AI-driven tools identify software vulnerabilities at a velocity that outstrips the capacity of existing tracking, remediation, and exploitation infrastructure. We are witnessing a decoupling where the sheer volume of discovered flaws bears little relation to the actual frequency of real-world exploitation.
The data backbone for this shift is stark. According to Epoch AI, June 2026 saw 1,500 high or critical CVEs reported from 21 major organizations—a 3.5x increase over the previous monthly record of approximately 430. This acceleration traces directly to the April 7, 2026, preview of Anthropic’s Claude Mythos. While the public release was withheld, the model’s internal performance—which included a 73% success rate on expert-level CTF challenges, the generation of 181 working Firefox exploits, and the completion of the UK AISI’s 32-step corporate network attack simulation—signaled that the barrier to entry for finding high-severity vulnerabilities had effectively collapsed. As Gabriel Bernadett-Shapiro, Distinguished AI Research Scientist at SentinelOne, noted: “We have to come to the reckoning that these tools are increasing the ability of people to find vulnerabilities in software.”
This explosion in discovery has overwhelmed the traditional gatekeepers of vulnerability management. The NIST National Vulnerability Database has been in a state of crisis, struggling to process a 263% increase in CVE submissions between 2020 and 2025. By April 2026, the system was so strained that NIST implemented a major overhaul, moving all pre-March 2026 CVEs to a “Not Scheduled” category to focus on current intake. The enrichment capacity of the NVD has been fundamentally broken by the sheer scale of AI-generated findings, forcing a shift toward risk-based prioritization that leaves a massive backlog of historical data unaddressed.
The operational reality for security teams is further complicated by the collapse of the exploitation window. In 2022, the mean time-to-exploit was approximately 32 days. By 2026, that window has shrunk to roughly 10 hours. AI now enables the generation of working proof-of-concept exploit code for published CVEs in 10 to 15 minutes at a cost of approximately $1 per attempt. Consequently, 32.1% of 2025 exploits appeared on or before the official CVE disclosure date, rendering traditional patch-management cycles obsolete.
Despite this velocity, there is a critical disconnect between discovery and impact. CISA’s Known Exploited Vulnerabilities (KEV) catalog, which tracks flaws actively used in the wild, has not seen a corresponding spike in additions. The volume of AI-discovered CVEs is decoupling from actual exploitation. A clear example is Azure OpenAI CVE-2026-45499, a critical SSRF vulnerability with a CVSS score of 9.9. Despite its high severity rating, the issue was fully mitigated server-side by Microsoft, requiring zero customer action and resulting in no real-world exploitation. It stands as a data point for the current environment: a critical vulnerability that exists in the database but carries no actual risk.
The industry and government response is now unfolding in three distinct layers of triage. At the government level, Executive Order 14409, signed June 2, 2026, established GOLD EAGLE, a Treasury-led clearinghouse involving the NSA and CISA to coordinate AI-driven scanning and prioritize remediation across federal and private sectors. At the industry level, Cisco’s PSIRT has moved to a biweekly, category-based disclosure cadence, bundling CVEs by CWE category rather than individual findings to manage the volume. Finally, the standards community is attempting to address the shift, as evidenced by the July 30, 2026, CVE Program virtual event, which focused on the challenges of scale, timing, and abstraction in an era of AI-enabled discovery.
The emergence of the Machine Control Protocol (MCP) security wave, which saw dozens of CVEs discovered between July 11 and July 21, 2026, including new prompt injection vectors, suggests that we are entering a period of rapid, iterative vulnerability discovery. This pace of discovery is forcing a transition from human-paced research to automated, high-frequency identification.
The current CVE model was designed for a world of human-paced research and manual verification. As AI continues to automate the identification of flaws, the industry is moving toward a state where the NVD and the broader disclosure ecosystem must either evolve to distinguish between theoretical vulnerabilities and actionable threats, or continue to be buried under the weight of its own discovery capacity.