For years, the security community’s worry about AI-powered vulnerability discovery was essentially theoretical: yes, these models will be able to find bugs at scale, yes, that will stress-test patching capacity, yes, it will change the economics of attack versus defense. Epoch.ai published analysis on July 2 that makes this concrete with actual numbers.
The methodology: track CVE submissions from 21 reputable organizations — Microsoft, Google, Apple, Adobe, Oracle, Cisco, IBM, Red Hat, Intel, AMD, NVIDIA, Qualcomm, Samsung, SAP, AWS, VMware, GitHub, Linux kernel maintainers, Mozilla, Apache, and OpenSSL — as a signal for genuine, carefully-vetted vulnerability disclosures rather than the noisier full CVE database. In June 2026, those organizations disclosed approximately 1,500 high- and critical-severity vulnerabilities. That is more than 3.5 times the previous monthly peak.
The timing lines up with Project Glasswing, Anthropic’s limited partnership program that deploys Mythos Preview for autonomous vulnerability discovery in critical infrastructure. Anthropic announced in April 2026 that Mythos could independently identify software flaws; Glasswing partners include Microsoft, Google, Apple, and AWS — four of the 21 organizations in Epoch’s tracked set. Project Glasswing reportedly found over 10,000 high- or critical-severity vulnerabilities. Most remain undisclosed.
What shows up in the public data is the tail of that iceberg. When a vulnerability is found, organizations decide what to patch, what to coordinate-disclose with other vendors, and what to sit on. The 1,500 public disclosures are presumably the subset that could be addressed and released without requiring coordination that hadn’t happened yet. The 10,000+ sits in a queue somewhere.
The Epoch authors are careful about causation. Two effects could explain the spike: increased discovery (AI finds more real bugs) and increased disclosure incentive (knowing that others are running similar tools creates pressure to disclose before being scooped). Both are probably operating. The fact that the spike is concentrated in high- and critical-severity categories — not low-severity noise — is some evidence that this reflects real discovery rather than just disclosure behavior shifts, since organizations don’t typically bump a low-severity bug to critical just because they want to look active.
The patching math is uncomfortable. If AI systems are finding vulnerabilities much faster than human researchers were, and the bottleneck for security is no longer finding bugs but patching them, then the current disclosure-and-patch cycle is increasingly mismatched to the discovery rate. A 3.5x monthly spike in serious CVEs is not something that existing security operations and patch management processes were designed to absorb. The organizations in the tracked set are the most capable in the world at handling this kind of volume; smaller companies patching against the CVE feed don’t have the same resources.
The flip side — and this is the standard argument, and it is also correct — is that better discovery is better than worse discovery. Finding 10,000 vulnerabilities and patching them is better than leaving them for attackers who would have found them anyway. The question is how long the defensive advantage lasts as offensive AI catches up to the same discovery capability.
What the Epoch analysis adds is a data point that was previously missing: the first visible evidence in public statistics that AI-powered vulnerability discovery is happening at scale and has a measurable footprint. People have been arguing about whether this transition was imminent or already underway. In June 2026, it shows up in the numbers.