AI-Driven Vulnerability Discovery Surges, Threatens Software Security

OpenClaw, a popular AI project, ranked 12th among those with the highest number of vulnerabilities discovered and published in Q2, with over 200 CVEs registered during the reporting period.

AI adoption and the surge in registered CVEs

The vulnerability landscape shifted sharply in Q2 2026. Kaspersky reports that the number of registered CVEs reached an unprecedented level, driven primarily by the widespread adoption of AI for both application development and vulnerability discovery. Kaspersky’s vulnerability knowledge base — which aggregates the CVE database, the Russian BDU database and GitHub Advisory (GHSA) — shows the monthly count of published vulnerabilities surging from 2022 through 2026, and a sharp jump in critical vulnerabilities (CVSS > 9.0) during Q2. The firm attributes the rise to AI’s ability to analyze massive amounts of previously unexamined code and to uncover entire classes of long‑standing, unnoticed flaws.

Researchers publishing exploits early: Windows examples

Q2 set a precedent for public exploit publication before CVE assignment or patching. A researcher known as Nightmare Eclipse (aka Chaotic Eclipse) published a set of “named” Windows vulnerabilities whose technical write‑ups and working exploits existed before CVE identifiers were issued. The list includes:

• BlueHammer — a local privilege escalation in Windows Defender via a TOCTOU race during signature database updates; a fully functional exploit was published.
• RedSun — a logical Windows Defender vulnerability with a working exploit allowing overwrite/restore of “cloud” marked files with elevated privileges.
• YellowKey — an exploit enabling bypass of BitLocker via the Windows Recovery Environment (WinRE); a fully functional exploit was published.
• GreenPlasma — a system object injection via the CTF loader for the Collaborative Translation Framework (CTFMON); original publication included exploit with limited functionality.
• RougePlanet — a Windows Defender TOCTOU vulnerability used to overwrite wermgr.exe with a malicious file.
• UnDefend — a Windows Defender vulnerability whose exploit causes denial-of-service and blocks updates.

Kaspersky notes that early publication gives attackers an advantage and that veteran Windows vulnerabilities — including CVE‑2018‑0802, CVE‑2017‑11882, CVE‑2017‑0199, CVE‑2023‑38831, CVE‑2025‑6218 and CVE‑2025‑8088 — remain commonly exploited. Despite this, Kaspersky’s telemetry found the number of Windows users who encountered exploits fell to an 18‑month low in Q2.

Dirty Frag and Linux page‑cache exploitation

Linux experienced its own concentrated wave of kernel exploitation in Q2. Kaspersky details a family of “Dirty Frag” vulnerabilities tied to the Linux caching and page cache mechanisms and the networking subsystem. The most actively exploited vulnerabilities include CVE‑2026‑31431 (Copy Fail), CVE‑2026‑43284 and CVE‑2026‑43500 (Dirty Frag variants affecting IPsec ESP and RxRPC), CVE‑2026‑46300 (Fragnesia), CVE‑2026‑31635 (DirtyDecrypt), CVE‑2026‑43494 (PinTheft) and CVE‑2026‑46331 (pedit COW). These let unprivileged users overwrite or modify the page cache and escalate to root — a serious risk for cloud and containerized environments. Kaspersky also records ongoing exploitation of older Linux bugs such as CVE‑2022‑0847 (Dirty Pipe), CVE‑2019‑13272, CVE‑2021‑22555 and CVE‑2023‑32233. Overall, the number of Linux users encountering exploits fell slightly versus Q1, but Kaspersky recommends rapid patching or disabling vulnerable kernel modules where patching is infeasible.

C2 frameworks, APTs, and AI‑tool exploitation

Kaspersky’s Q2 analysis finds APT groups increasingly exploit new vulnerabilities immediately on publication and continue to leverage zero‑days. The report highlights Langflow as a notable case — “one of the first cases of an APT group exploiting AI technology,” with many AI integrations described as having security blind spots. On the C2 side, open sources show Sliver, Havoc, AdaptixC2 and Metasploit among the most used frameworks. Vulnerabilities tied to APT activity that interacted with C2 agents include CVE‑2026‑35273 (Oracle PeopleSoft PeopleTools SSRF), CVE‑2023‑46604 (Apache ActiveMQ insecure deserialization), CVE‑2024‑12356 and CVE‑2026‑1731 (BeyondTrust command injection), CVE‑2023‑36884 (Windows Search bypassing mark‑of‑the‑web), CVE‑2025‑53770 (SharePoint insecure deserialization) and WinRAR directory‑traversal bugs CVE‑2025‑8088 and CVE‑2025‑6218. Kaspersky emphasizes these were used for initial access and privilege escalation to stage C2 deployments.

What this means for technologists, procurement leaders, and end users

• Technologists and security teams: prioritize rapid patching for kernel modules and AI‑integrated components, enable real‑time monitoring and consider disabling vulnerable kernel modules where immediate patching is impossible — all steps Kaspersky recommends for Q2 threats.
• Procurement and platform owners: assess AI tools and plugins for inadequate access control, improper authentication, and injection risks; the Kaspersky knowledge base shows these classes are prominent in AI/LLM‑related vulnerabilities (including undetermined CWE classifications).
• End users and administrators: beware polished, AI‑generated vulnerability write‑ups — Kaspersky reports fake Exchange and SharePoint research that described nonexistent problems and in some cases bundled unrelated malware, used to distract researchers.

The quarter’s paradox is clear: AI has accelerated discovery of software flaws while also contributing new attack surfaces in AI tools themselves. Kaspersky’s advice is to pair traditional patch management with real‑time monitoring, stricter access controls, and preventive endpoint measures — a capability Kaspersky says its Kaspersky Next platform provides. The immediate question left by Q2 is operational: can organizations move as fast to harden AI integrations and kernel modules as automated tools move to find flaws?

Original report