Beazley says agentic AI drove 36% rise in newly disclosed vulnerabilities
Beazley Security’s Q2 2026 threat report says AI tools pushed a sharp rise in newly disclosed vulnerabilities, while ransomware intrusions still mostly began with stolen credentials.
Beazley Security said on Aug. 18 that agentic AI helped drive a 36% quarter-over-quarter increase in newly disclosed vulnerabilities in its Quarterly Threat Report for Q2 2026.
The report said disclosure volume had typically moved within a 10% band from one quarter to the next. Beazley said that pattern changed in 2026, with disclosures rising 18.5% in Q1 and 36% in Q2.
The company said the increase reflects the rapid operationalization of agentic AI across vulnerability research programs. It also said the number of vulnerabilities confirmed as actively exploited and added to the Cybersecurity and Infrastructure Security Agency’s Known Exploited Vulnerabilities catalog rose 10% over the same period.
The report pairs that increase with a more familiar attacker pattern. Beazley Security said compromised credentials used against internet-facing virtual private networks and remote desktop services accounted for 67% of ransomware intrusions it investigated, down from 74% in Q1 but still the largest share. It also said public ransomware leak-site postings fell slightly to 2,268, while remaining nearly 60% above Q2 2025.
Beazley said business email compromise remained among the most common incident types. The report also said attackers are increasingly abusing Microsoft’s device code authentication flow to capture session tokens.
In a statement cited in the release, Beazley Security CEO Alton Kizziah said AI has made the security industry’s job noisier without making the attacker’s job fundamentally different. He also said AI-assisted attacks are gaining in both frequency and effectiveness.
The report placed its findings in a broader industry context. It cited changes and incidents including NIST no longer enriching every new CVE, HackerOne’s Internet Bug Bounty pausing submissions, Pwn2Own issuing applicant rejections for the first time, Cisco restructuring its disclosure model, the TanStack package compromise tied to TeamPCP, Sysdig’s documentation of JADEPUFFER as a ransomware campaign driven end to end by a large language model, and law enforcement’s Operation ENDGAME takedown followed by a new StealC version within four days.
Beazley’s report said organizations should consider AI assessments to track what AI capabilities are in use, how they are used and what controls are needed to manage them.
Beazley Security said its Q2 2026 report was released Aug. 18.
Sources
Editorial note: This article was prepared with AI assistance and includes source attribution for reader review.
View the Web StoryA visual, tap-through version of this reportThis is the AMP version of the article. Read the full article on Atlas Watchline.