TechAILatest.newsTrackers
Cybersecurity / AI security tools

Microsoft MDASH turns 26,502 AI findings into 59 Linux kernel bug reports and 33 CVEs

Microsoft says its MDASH AI system filed 59 Linux kernel bug reports, led to 33 CVEs and 8 working root exploits, but fixes took about five months to merge.

By Gürkan E.Published

Image Microsoft published with the announcement covered in this storyTech AI Latest
Image published by Microsoft with the announcement. Image: Microsoft

Microsoft's FORGE Lab said on 11 November 2026 that its AI scanning system MDASH produced 26,502 raw findings in selected Linux kernel subsystems, which researchers narrowed to 621 cases with proof of concept and 59 public bug reports. The reports led to 33 CVEs, 24 rated High or above, and 8 working root exploits. Microsoft's main lesson: finding bugs was fast, but validating and fixing them took months of human work.

Key facts

  • Published: 11 November 2026 (Microsoft Security Blog)
  • Funnel: 26,502 raw findings, 11,769 candidates, 621 cases with proof of concept, 59 upstream reports
  • Outcome: 33 CVEs (24 rated CVSS 7.0 or higher), 8 confirmed root privilege escalation exploits, 38 fixes merged into mainline
  • Oldest bug: a cls_u32 flaw present for at least 21 years; average age 8.1 years
  • Models: Sonnet 4.6, GPT-5.4 and GPT-5.5 for scanning; GPT-5.5, Opus 4.6 and GPT-5.5-cyber for reproduction
  • Not stated: the cost of the scans in compute or API spend

What Microsoft published

Microsoft posted Lessons from 59 Linux Kernel Bug Reports on its security blog on 11 November 2026. The post, from the company's FORGE Lab and researcher Cen Zhang, describes a three-month project in which MDASH, Microsoft's AI system for vulnerability research, scanned a subset of Linux kernel subsystems, among them networking, filesystems and security modules, from late May to mid-June 2026. One full-time researcher and one intern led the work.

The scans produced 26,502 raw findings in about two weeks on two servers. Filtering cut them to 11,769 candidates, and 621 cases came with proof-of-concept evidence reproduced in QEMU virtual machines. Of those, 59 became public upstream reports, the first sent on 27 June 2026. Microsoft says the reports have led to 33 CVEs, 24 of them rated High or above, and 38 fixes merged into the mainline kernel. Its researchers confirmed working local root exploits for 8 of the CVEs. Microsoft lists the models it used: Sonnet 4.6, GPT-5.4 and GPT-5.5 for scanning, and GPT-5.5, Opus 4.6 and GPT-5.5-cyber for reproducing bugs.

Why most findings were never filed

The gap between 621 proven cases and 59 reports is the most useful part of the post. Microsoft says 266 cases shared a root cause with another case, so one patch fixed several; 163 were dropped after review, mostly because their security impact was low; and 133 had already been reported or fixed upstream. The bugs that remained were old: on average 8.1 years at the time of reporting, and one shift-out-of-bounds flaw in the cls_u32 traffic classifier, CVE-2026-74544, had been in the kernel for at least 21 years.

Microsoft's detailed example is CVE-2026-64563, a one-line state reset bug in rhashtable, the kernel's shared resizable hash table. A 2018 change stopped clearing a saved iterator pointer on one resume path, so a paused walk could follow a freed entry after a resize. Netlink socket diagnostics, which unprivileged users can reach, made it exploitable for local privilege escalation, including on Android with SELinux enforcing. Microsoft notes that Yuan Tan and Frank Wu of NebuSec found and exploited the same bug independently.

The bottleneck is people, not models

The post's main argument is that discovery now outruns repair. Reporting the 59 bugs took more than four person-months, and the 38 merged fixes took about five months. Half were accepted on first submission and the rest needed one to four revisions; the median time from first public patch to a mainline release was 41 days, and the longest was 108. Microsoft says more tokens can find more bugs but do not by themselves make a system secure, and that maintainers can be swamped by volume, so reports should focus on high severity issues.

The numbers sit next to OpenAI's report of 948 validated kernel privilege escalation flaws from a single run on 4 November, and Anthropic's Claude exploit chains against Chrome and Linux in October. Microsoft's account is the first of these to report how many findings survived human triage and how long fixes took to land.

What to watch next

Microsoft says it is feeding the lessons back into MDASH and that the project covered only a small part of the kernel, so more reports are likely. For Linux users the practical step is to track distribution kernel updates that include the listed CVEs, especially on shared hosts where a local root bug turns any foothold into full control. Further coverage of AI-driven vulnerability research is in the AI cybersecurity section.

What changed

BeforeMicrosoft had described MDASH results on Windows; no account of a large AI-assisted Linux kernel campaign from Microsoft with triage and fix times
AfterMicrosoft reports 59 Linux kernel bug reports, 33 CVEs, 8 root exploits and 38 merged fixes from 26,502 AI findings, with fixes taking about five months
Who it affectsLinux kernel maintainers, distribution vendors, Android device makers and operators of shared Linux hosts

Questions people ask

How many Linux kernel vulnerabilities did Microsoft's AI find?
Microsoft filed 59 public bug reports from 621 cases with proof of concept, which led to 33 CVEs. The 26,502 raw findings were mostly duplicates, low impact or already known.
Which AI models did Microsoft use?
Sonnet 4.6, GPT-5.4 and GPT-5.5 for scanning, and GPT-5.5, Opus 4.6 and GPT-5.5-cyber for reproducing bugs, according to the Microsoft post.
Are the bugs fixed?
Microsoft says fixes for 38 of the 59 reports have been merged into the mainline kernel. Users get them when their distribution ships an updated kernel.

Sources

  1. Lessons from 59 Linux Kernel Bug Reports (Microsoft, 11 Nov 2026)
  2. AI found Linux kernel bugs that remained undetected for decades (CSO Online, 12 Nov 2026)

Reference pages

Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.

Comments

Loading comments.

No links, no insults; comments that break the rules are not published. 0/1000

Related stories

All cybersecurity