AINEWS
AI Is Finding More Security Flaws Than Defenders Can Handle
A 2026 CVE forecast of roughly 66,000 vulnerabilities puts triage—not discovery—as the binding constraint on enterprise defense.
A 2026 CVE forecast of roughly 66,000 vulnerabilities puts triage—not discovery—as the binding constraint on enterprise defense.
Security operations centers used to measure their week in patches shipped. By late June 2026, the more honest unit of measurement is what is still sitting in the triage queue, and that queue is filling faster than any patch window can drain. The driver is not a single dramatic zero-day leak but a quiet acceleration in how quickly AI tooling finds previously hidden vulnerabilities in code that defenders already shipped years ago.
Help Net Security's first 2026 forecast puts the year's CVE total on track for roughly 66,000 disclosures (Help Net Security). That number is a forecast, not a measured count, and it sits on top of a base that was already climbing. The point is not the absolute figure but the slope: discovery capacity is now rising faster than remediation capacity, and that gap is the binding constraint for enterprise defense.
The acceleration has at least three named drivers. Anthropic's research update on Project Glasswing describes work on automated vulnerability discovery, and a separate Register report in April covered Anthropic's earlier claim that its Mythos model could find and exploit zero-days in a controlled setting (Anthropic, The Register). OpenAI has moved into the same lane with Daybreak, a security-tooling initiative that recasts a model lab as a defender-side vendor (OpenAI). Chainguard's Athena coalition is the industry-coordination response: a shared effort to harden the software supply chain against AI-driven bug finding (Chainguard). Capability claims from these vendors are still company-reported in most cases, and independent reproduction of mass exploitation is not yet on the public record.
The result is a defender landscape where a single week can deliver a stalled patch on a long-running SharePoint zero-day under active attack, an Amazon Q flaw that lets a booby-trapped Git repository execute commands and steal cloud credentials, and a Russian-aligned Signal phishing campaign that hijacks support chats (The Register). That is one curated week's worth of items. Multiply it across the year and the workflow problem becomes obvious: more findings, more sources of findings, and the same human-hours to sort, validate, prioritize, and patch.
The constructive read is that the same AI tooling accelerating discovery can also accelerate triage, if defenders restructure their workflow around it. AI-assisted prioritization, automated patch drafting, and continuous validation against known-good baselines all sit inside the same technology curve. None of them is a free lunch. Adoption still depends on staffing, integration with existing ticketing systems, and the willingness to trust a model's ranking over an analyst's gut.
The honest caveat is that more discovery does not automatically mean more security. False positives carry cost, and security teams remain chronically under-resourced. The teams coping best are not the ones with the largest headcount or the loudest AI vendor. They are the ones who rebuilt their triage pipeline around the assumption that the queue will keep growing, and who treat prioritization rather than discovery as the binding constraint.
What to watch next: whether any major enterprise publishes patch-cycle time as a primary operational metric, and whether the named vendor coalitions publish independent reproductions of their discovery claims rather than internal benchmarks. Either signal would tell the reader whether the speed gap is closing, or just widening faster.