The number of software vulnerabilities disclosed each month roughly doubled between January and August 2026, and attackers exploited more flaws in the first eight months of the year than in all of 2025, Google Threat Intelligence Group said in a report published Wednesday. The authors, Robin Grunewald, Supriya Mazumdar and Kelli Vanderlee, attribute much of the change to AI, both in how bugs are found and in how quickly attackers turn fixes into working exploits.

What Happened

GTIG counted 5,045 disclosures in January and 10,740 in August. Flaws it rates as high risk grew faster, by 167%, from 131 to 350 a month, though they remain about 3% of the total. Part of the volume comes from a few sources: the Linux kernel alone generated about 5,000 CVEs over the period without a single in-the-wild zero-day among them, and Oracle's quarterly patch release and Linux network-driver advisories accounted for 128 high-risk flaws in August.

Exploitation rose too. GTIG observed 141 vulnerabilities exploited from January through August, compared with 127 in all of 2025, lifting the monthly average from 10.5 to 18. Zero-days averaged 11 a month against 8 last year, with an unusual spike to 22 in August, and made up 62% of exploited flaws. Even so, only about one in 431 disclosed vulnerabilities, or 0.23%, has been seen exploited.

RelatedNew curl Release Fixes 24 Flaws, Including a High-Severity Bug Reported With Anthropic’s Claude

Why It Matters

The report's central finding is about speed rather than novelty. GTIG says the growth in exploitation is driven mainly by faster weaponization of n-days, flaws that already have a patch, with attackers using large language models to study patches, advisories and proof-of-concept code. Speaking to The Record, Vanderlee said comparing product versions with AI tools can be a more efficient path for attackers than hunting for zero-days.

The BeyondTrust case shows the timeline. CVE-2026-1731, a command-injection flaw in BeyondTrust's remote support products found by an autonomous research agent from Hacktron AI, was exploited by one threat cluster within four days of disclosure and by five more within seven days.

AI-found bugs also look different. Half of the vulnerabilities discovered with AI lead to remote code execution, compared with 26% of those found by other means, and AI discovery skews toward medium-risk rather than low-risk issues. GTIG tracked 2,076 AI-related CVEs from January 2025 through August 2026. Agent orchestration frameworks such as Flowise, Langflow and LangChain accounted for 782 of this year's AI flaws, often through code-execution features that can be reached by prompt injection.

Where Attacks Land

Edge and security appliances made up 14% of exploited vulnerabilities this year, and more than 65% of the edge flaws exploited carried high or critical risk ratings. Enterprise directory and collaboration software accounted for another 11%. The Record cited a CISA projection of about 96,000 new CVEs by the end of 2026, with more than 67,000 published through September.

What to Watch

GTIG recommends moving from patching everything to triage driven by threat intelligence, compressing response times for n-days, sandboxing autonomous agent workloads, requiring authentication on inference gateways and running AI code review before software ships. It expects discovery and exploitation rates to keep climbing in the short to medium term and warns that defenders have limited time before attackers scale up agent-driven exploit tooling. The next signals are whether the August zero-day spike holds and whether CISA's year-end CVE count lands near its projection.