Skip to content
News Research

AI vulnerability hunters drove a 3.5x surge in high-severity CVE disclosures in June 2026, per Epoch AI data

· by Pondero Newsdesk

The short version

Epoch AI tracked around 1,500 high- and critical-severity CVEs from 21 major organizations in June 2026, more than 3.5 times the monthly record before Anthropic released Claude Mythos Preview.

AI vulnerability hunters drove a 3.5x surge in high-severity CVE disclosures in June 2026, per Epoch AI data

Around 1,500 high- and critical-severity CVEs were disclosed by 21 major software organizations in June 2026, a figure more than 3.5 times the monthly record set before Anthropic released Claude Mythos Preview, per Epoch AI's July 2 data insight. The number gives the first empirical dataset quantifying how AI-driven vulnerability discovery is reshaping the security disclosure pipeline.

What the data shows

Epoch AI researcher Luke Emberson published the analysis on July 2, tracking Common Vulnerabilities and Exposures (CVEs) from 21 named organizations: Microsoft, Google, Apple, Adobe, Oracle, Cisco, IBM, Red Hat, Intel, AMD, NVIDIA, Qualcomm, Samsung, SAP, Amazon (AWS), VMware, GitHub, Linux, Mozilla, Apache, and OpenSSL.

The organization-level filter matters. Epoch excluded lower-credibility submissions to focus on disclosures that carry real downstream weight for operators. The June total of roughly 1,500 high- and critical-severity CVEs from that group exceeded the previous monthly record by more than 3.5 times, with the climb beginning in April 2026 when Anthropic publicly announced Claude Mythos Preview.

Anthropic described Mythos Preview as capable of autonomous software vulnerability discovery and exploitation. Per Anthropic's own announcement, Project Glasswing partners, including Microsoft, Google, Apple, and AWS, had already been using the model ahead of public release. Project Glasswing claims more than 10,000 high- or critical-severity vulnerabilities found to date, per Anthropic, with many not yet individually disclosed. OpenAI's Daybreak program is running a parallel effort, per The Decoder.

Why it matters

The data shifts a key security assumption: AI models are no longer helping human researchers find bugs faster at the margins. They are generating enough distinct findings to alter the monthly disclosure cadence for major vendors by a multiple.

That has two practical consequences for anyone running software infrastructure. First, the pace of patch releases from the named vendors is accelerating, which compresses the window between disclosure and exploitation. Security teams that rely on monthly or quarterly patch cycles will find that window shortening. Second, the bottleneck has moved. Epoch AI notes the increase may also reflect heightened interest in discovering bugs rather than AI capability alone, but either way the triage and response workload is growing faster than the tooling most teams use to handle it.

The 10,000-vulnerability claim from Project Glasswing is Anthropic's own figure and not independently verified. Epoch's tracked CVEs represent only the publicly disclosed subset, which makes the 1,500 figure a floor, not a ceiling.

Context

Anthropic announced Mythos Preview in April 2026 through its red-team site, framing it as a system that could find and exploit vulnerabilities faster than developers can patch them. The Glasswing program, which pairs Anthropic engineers with major technology vendors, predates that public release. OpenAI's comparable Daybreak program launched shortly after.

The surge in disclosures represents security vendors burning through a backlog of AI-identified findings, not necessarily new attack surface appearing overnight. Several organizations in Epoch's list operate under coordinated disclosure policies that hold findings for weeks or months before publication, so the June peak likely captures bugs found in prior months.

APNIC noted in a July 1 blog post that the spike raises questions about whether network operators and enterprise security teams have the triage capacity to process the volume, given that each high-severity CVE requires evaluation and typically a remediation decision.

What to watch next

Two signals are worth tracking in the next 60 to 90 days. First, whether CISA issues guidance on AI-accelerated CVE triage, which would signal the U.S. government views the disclosure volume as a systemic risk rather than a one-time spike. Second, whether the disclosure rate plateaus or continues climbing as additional AI vulnerability-hunting programs ramp up. If it continues, the market for AI-based prioritization tools that help teams sort critical from merely high-severity findings will likely see new entrants.

Sources