Finding 59 kernel bugs is a solid result. What makes it interesting is that some had sat in production code for more than two decades, and the AI part was the fast part. Validation, patching, and working with maintainers stayed stubbornly human.
From 26,502 findings to 59 reports
The work ran under the codename MDASH inside Microsoft's FORGE Lab. Over roughly two weeks, the team scanned selected kernel subsystems covering networking, filesystems, and security modules and generated 26,502 potential findings. Automated screening cut that to 11,769 candidates. AI agents then tried to reproduce the remaining issues using QEMU, the open-source machine emulator, which produced 621 cases with proof-of-concept evidence.
That's where people took over. Researchers worked through those 621 cases to pin down root causes, merge duplicates, and prepare upstream submissions. Of those, 59 became public upstream reports. Another 266 shared root causes with other findings and could be addressed by the same patch. A further 163 were dropped after review, mostly for limited security impact, and 133 had already been reported or fixed upstream.
Microsoft filed its first public report on June 27, 2026. By September 1, 55 of the 59 had been posted to public mailing lists. The remaining four were reported privately during that window and routed to public lists by mid-October. Of the 33 CVEs assigned, 24 carry a High or higher rating, meaning CVSS v3.1 scores of at least 7.0.
Eight of them came with working exploits
Microsoft built working local privilege-escalation exploits for eight of the bugs. They touch IPv4 and IPv6 networking, IPsec routing, encrypted-key handling, and the resizable hash-table implementation. Some demonstrations needed non-default settings or disabled exploit mitigations, and that caveat matters when you assess real-world exposure.
CVE-2026-64563 is the one I'd look at first. It sits in rhashtable, a shared resizable hash-table library used across several kernel subsystems. Microsoft traced the flaw to a 2018 code change that let a table-walk operation keep a pointer while paused. Under specific conditions involving table resizing and the removal of an entry, the kernel could later access freed memory through that stale pointer. The researchers found an exploitable path through Netlink socket diagnostics, which unprivileged users can reach. Manipulating memory could turn the bug into a constrained kernel memory-write primitive that, combined with other techniques, enables local privilege escalation.
Microsoft said the bug worked on standard Linux configurations without special kernel options or user namespaces. It also affected Android devices running SELinux in enforcing mode. That's the detail that moves this from a niche configuration problem to something much wider.
The bugs were old. On average, they had been present for 8.1 years before Microsoft reported them publicly. CVE-2026-74544 had been in the kernel for at least 21 years.
The patch pipeline is the real constraint
AI agents helped identify candidates, generate proof-of-concept demonstrations, and draft initial patches. But Microsoft was clear that human investigation still did the heavy lifting on validation and maintainer communication.
"Upstream Linux typically requires vulnerability reports to come with a working patch. Agents helped draft initial fixes, but human validation, subsystem understanding, and iterating with maintainers during review remained essential," the researchers wrote.
Fixes for 38 of the 59 reports had merged into upstream mainline at publication. Half were accepted on the first submission, while the others needed between one and four revisions. The median time from first public patch submission to mainline release was 41 days, and the longest took 108 days. Work on the remaining reports was ongoing.
The scans took about two weeks. Turning those findings into validated reports and patches took more than four person-months. The researchers put it bluntly: "Finding bugs and fixing them are drastically out of balance. For complex infrastructure like the Linux kernel, keeping humans in the loop is still necessary, making human effort the bottleneck."
Microsoft acknowledged the project covered only a small fraction of the kernel, and further tuning could surface more bugs. That's the part I'd watch. Scanner throughput is no longer the limiting factor. Maintainer bandwidth is, and more AI-generated reports will only make that gap wider. As the team put it, "More tokens may yield more bugs, but not a secure system."