Security patches need patches now that attackers have AI

Broken fixes rise in 2026

Terra Security, an AI penetration-testing startup, reviewed published CVE records and GitHub-reviewed advisories for fixes that failed. Researchers said patch bypasses had remained relatively uncommon for six years before rising sharply in 2026.

The proportion of GitHub security fixes followed by an advisory reporting failure within 90 days was about 0.2 percent from 2019 through 2025. In 2026, Terra found it had climbed to 2.3 percent.

A Windows fix fell short

Russian government cyberspies exploited a Windows Shell security-bypass flaw, CVE-2026-21510, at least a month before Microsoft patched it in February. The patch was incomplete, according to the report, leaving another vulnerability.

The remaining flaw, CVE-2026-32202, allowed zero-click authentication coercion and was exploited before Microsoft fixed it in April. The sequence illustrates how a patch can leave another route open to attackers.

Some vulnerability types stand out

Terra found that server-side request forgery fixes and prototype-pollution patches were most likely to break. Compared with their share of all CVEs, prototype-pollution flaws appeared about 11 times as often among broken-fix CVEs, and server-side request forgery flaws appeared 4.6 times as often.

The researchers warned that a blocklist may miss an encoding of harmful input. They also reported that 3.2 percent of kernel CVEs were introduced by fixes for other CVEs, rising to about 10 percent in busy areas including fs/smb and net/sched.

Patch managers urged to stay alert

Terra’s penetration testers advised administrators to treat fixes for exploited vulnerabilities cautiously. They said attackers may examine a vendor’s patch to find a way around it. The findings point to the need to check whether a fix fully addresses the underlying issue. A vendor’s update may be followed by a bypass and then another fix, the researchers said.