In 2025, the CVE ecosystem hit a record: 48,185 published vulnerabilities — a 20.6% increase over 2024's 39,962 [1]. That's 131 new CVEs every single day. On February 26th alone, 793 CVEs were published in a single day [1].
The projections for 2026 are worse. FIRST.org's statistical forecast puts the median at approximately 59,000 CVEs — the first year to cross the 50,000 threshold. Their 90% confidence interval stretches from 30,000 to nearly 118,000 [2]. Infosecurity Magazine called it "record-breaking" before the year even started [3].
But here's what most vulnerability management teams are missing: this isn't about developers writing worse code. The volume explosion isn't driven by new bugs being introduced. It's driven by old bugs being found.
The bugs were always there. AI is just finding them now — all at once.
In November 2024, Google announced the results of adding LLM-powered fuzz target generation to their OSS-Fuzz platform. The results: 26 new vulnerabilities discovered across open-source projects, improved code coverage across 272 C/C++ projects, and over 370,000 lines of new fuzz target code generated [4]. The most notable find: an exploitable out-of-bounds read/write flaw in OpenSSL (CVE-2024-9143) that had been present in the codebase for 20 years [4]. Two decades of human review, traditional fuzzing, and security audits — and it took an AI to find it.
Google's Big Sleep project — a collaboration between Project Zero and DeepMind — went further. In July 2025, Big Sleep discovered CVE-2025-6965, a memory corruption flaw in SQLite affecting all versions prior to 3.50.2. Google stated the vulnerability was "known only to threat actors and was at risk of being exploited" [5]. This wasn't a theoretical research exercise. An AI agent directly disrupted what appeared to be an active exploitation attempt.
Then there's XBOW, an autonomous offensive security platform. In under 90 days, XBOW reached the #1 position on HackerOne's U.S. leaderboard. The numbers: 285 total vulnerabilities reported, 130+ confirmed fixes, 22 confirmed CVEs from Docker Hub images. In a head-to-head comparison, a seasoned human pentester took 40 hours to work through 104 real-world scenarios. XBOW completed them in 28 minutes [6].
Trend Micro launched AESIR in mid-2025 — an AI system combining threat intelligence (MIMIR) with zero-day discovery via agentic code analysis (FENRIR). Since launch, AESIR has uncovered 21 critical CVEs across NVIDIA, Tencent, MLflow, and MCP tooling platforms [7].
Praetorian demonstrated a multi-agent AI pipeline for CVE research that compresses 4-8 hours of engineer work into under 30 minutes, producing production-ready detection templates overnight [8].
And then there's Anthropic's Mythos. In a single run, Mythos developed 10 separate zero-days to full control-flow hijack and found hundreds of crashes. The bugs it surfaced include a 27-year-old remote DoS in OpenBSD, a 16-year-old RCE in FFmpeg, a Linux privilege escalation requiring four chained exploits, and a remote root on FreeBSD's NFS server that required splitting ROP gadgets across multiple network packets. They sent Firefox 112 bug reports — none were rejected [10]. As Plaid's Head of Platform Security put it: this is "the air raid siren for the vulnerability cataclysm." And critically, Mythos is mostly a product of general frontier model improvement, not security-specific engineering — which means similar capabilities will soon be widely available.
Every year for the past decade, someone has written a "CVEs hit a new record" article. The industry has gotten used to gradual increases — a few thousand more each year, manageable with headcount growth and better tooling.
What's happening now is different. These AI-powered research tools are not incremental improvements to existing methods. They represent a fundamentally new discovery capability operating against the same codebase that's been running in production for years. The backlog of latent vulnerabilities in foundational software — the libraries, kernels, runtimes, and packages your cloud infrastructure depends on — is being systematically surfaced.
At RSAC 2026, Alex Stamos put it bluntly: "The exploit discovery has gone exponential." He added: "Foundation model companies are sitting on thousands of bugs discovered through AI-assisted analysis" [9]. Kevin Mandia called it "a perfect storm for offense over the next year or two" [9].
Your scanner output is about to double. Not because your environment changed — because the discovery rate did.
The scanners themselves don't get smarter. They detect what's in their signature databases, and those databases are about to get a lot bigger. Every one of those 59,000+ CVEs projected for 2026 that touches something in your environment becomes a finding. Your triage queue grows at the rate of global discovery, not at the rate of your own change.
The question every vulnerability management leader needs to answer: is your program designed for the volume that's coming, or for the volume you had?
[1] Gamblin, J. (2026, January 1). "2025 CVE Data Review." https://jerrygamblin.com/2026/01/01/2025-cve-data-review/
[2] FIRST.org. (2026, February 11). "Vulnerability Forecast for 2026." https://www.first.org/blog/20260211-vulnerability-forecast-2026
[3] Infosecurity Magazine. (2026). "FIRST Forecasts Record-Breaking 50,000+ CVEs." https://www.infosecurity-magazine.com/news/first-forecasts-record-50000-cve/
[4] Google Security Blog. (2024, November). "Leveling Up Fuzzing: Finding More Vulnerabilities with AI." https://security.googleblog.com/2024/11/leveling-up-fuzzing-finding-more.html
[5] The Hacker News. (2025, July). "Google AI 'Big Sleep' Stops Exploitation of Critical SQLite Vulnerability." https://thehackernews.com/2025/07/google-ai-big-sleep-stops-exploitation.html
[6] Dark Reading. "An AI-Driven Pen Tester Became a Top Bug Hunter on HackerOne." https://www.darkreading.com/vulnerabilities-threats/ai-based-pen-tester-top-bug-hunter-hackerone
[7] Trend Micro. (2026, January). "Introducing AESIR." https://www.trendmicro.com/en_us/research/26/a/aesir.html
[8] Praetorian. "AI-Powered CVE Research: Winning the Race Against Emerging Vulnerabilities." https://www.praetorian.com/blog/ai-powered-cve-research-winning-the-race-against-emerging-vulnerabilities/
[9] CyberScoop. (2026). "Security leaders say the next two years are going to be 'insane.'" https://cyberscoop.com/ai-cyberattacks-two-years-insane-vulnerabilities-kevin-mandia-alex-stamos-morgan-adamski-rsac-2026/
[10] Anthropic. (2026). "Assessing Claude Mythos Preview's cybersecurity capabilities." https://red.anthropic.com/