AI Security Roundup: Two Microsoft Defender Zero-Days Exploited, Verizon DBIR Finds Patching in Crisis, and AI Is Now Finding Hundreds of Chrome Bugs

Three security stories worth your attention from the past 48 hours: Microsoft shipped out-of-band Defender patches for two actively exploited vulnerabilities, the annual Verizon DBIR confirmed that vulnerability exploitation has overtaken stolen credentials as the leading breach vector for the first time in the report’s 19-year history, and a SecurityWeek analysis points to AI as the driver behind a dramatic surge in Google-reported Chrome vulnerabilities.

Microsoft Patches Two Actively Exploited Defender Zero-Days (CVE-2026-41091 and CVE-2026-45498)

On May 21, 2026, Microsoft pushed out-of-band security updates for two Windows Defender vulnerabilities that were already being exploited in the wild. The first, CVE-2026-41091 (CVSS 7.8), is a link-following privilege escalation flaw in the Microsoft Malware Protection Engine that lets a low-privileged attacker gain full SYSTEM-level control. The second, CVE-2026-45498 (CVSS 4.0), allows an attacker to throw Defender into a denial-of-service state, effectively disabling protection without alerting the user. Both vulnerabilities are variants of the BlueHammer exploit chain that a researcher dropped publicly last month. The fix ships in Microsoft Defender Antimalware Platform version 4.18.26040.7, and CISA added both CVEs to its Known Exploited Vulnerabilities catalog with a federal patching deadline of June 3. If you have automatic Defender updates enabled you’re likely already covered, but it’s worth verifying: open Windows Security → Virus & threat protection → Protection updates → Check for updates, then confirm the Antimalware Client Version is 4.18.26040.7 or later. (Source: Bleeping Computer | SecurityWeek)

Verizon 2026 DBIR: Vulnerability Exploitation Overtakes Stolen Credentials as the Top Breach Vector

The 2026 Verizon Data Breach Investigations Report, released May 22, contains a significant milestone: for the first time in the report’s 19-year history, vulnerability exploitation edged out credential theft as the leading initial access method, accounting for 31% of documented incidents. The data covers November 2024 through October 2025, so recent AI-acceleration isn’t yet baked in — this shift reflects organizations simply falling behind on patching. The median time to patch climbed to 43 days (up from 32), only 26% of CISA KEV-listed critical vulnerabilities were fully remediated, and ransomware now appears in 48% of all breaches. The practical read here: the KEV catalog should be driving your patch prioritization, not CVSS score alone — and if you’re still doing quarterly patch cycles, that gap is actively being exploited. (Source: Verizon 2026 DBIR | CPO Magazine)

Google’s Chrome Vulnerability Surge Points to AI-Assisted Discovery at Scale

SecurityWeek reported May 21 that more than 200 Chrome vulnerabilities patched in recent releases were tagged as “reported by Google” — a number that jumped from a handful in March to 100 in the May 5 advisory alone. Google hasn’t confirmed which specific AI tooling is responsible, but the timing lines up with broader industry trends: the company noted when it recently cut bug bounties that AI and automation have helped its teams move at “an unprecedented rate.” Google has its own internal tools (Big Sleep, CodeMender) and is one of roughly 50 organizations with access to Anthropic’s Claude Mythos model. For Microsoft security practitioners, this is the clearest signal yet that AI-assisted vulnerability discovery is going to keep accelerating patch volumes across all major platforms — plan your Secure Score and compliance workflows accordingly. (Source: SecurityWeek)

What to Watch

The BlueHammer exploit family still has an unpatched variant (CVE-2026-3220) that Microsoft flagged as actively exploited during the April Patch Tuesday — keep an eye on whether an additional out-of-band fix lands for that one. Meanwhile, if your organization is running any agentic AI workflows on Azure SRE Agents or Claude Code, the broader sandbox-bypass research surfacing this month is worth a read.