Microsoft has published the results of an AI-assisted hunt for security holes in the Linux kernel: 59 bug reports sent to the project, 33 of them now with CVE numbers, and eight bugs the team turned into working exploits that give a normal local user full root control. Mainline Linux has already merged fixes for 38 of the reports.
Inside Microsoft’s FORGE Lab, a researcher and an intern ran the tool, called MDASH, across chosen kernel areas between late May and mid-June 2026. They chose networking, filesystems and security modules, because those parts run without special hardware and can be tested in a QEMU virtual machine.
Microsoft’s own account is candid about how little of the raw output survived. The scans, run on two servers over about a fortnight, flagged 26,502 possible problems. Screening left 11,769, and AI agents then tried to reproduce each one. That gave 621 cases with a demonstrated crash or exploit, and only 59 turned into public reports. Another 266 were the same underlying fault as a case already counted, 163 were judged too harmless to report, and 133 were already known upstream.
Age is the striking part. One report, CVE-2026-74544 in the cls_u32 traffic classifier, covers a flaw that had sat in the code for 21 years at the least. The reported bugs averaged 8.1 years in the kernel, and three of them were 20 years old or more. Microsoft rates 24 of the 33 CVEs High or above.
The eight exploits hit IPv6, IPsec, route handling, a GTP tunnel driver and the kernel’s key storage. Microsoft spends the most time on CVE-2026-64563, a single missing line in a shared hash-table routine. It worked without unusual kernel settings and also on Android with SELinux enforcing. A researcher at NebuSec found the same bug independently and shared working exploits. For some of the other seven, Microsoft says it had to use non-default settings or switch off protections, and everything was demonstrated only in lab conditions.
Most of the scanning was done by Anthropic’s Sonnet 4.6, which accounted for 71.1% of the tokens, with OpenAI’s GPT-5.4 and GPT-5.5 covering the remainder; GPT-5.5 also did most of the reproduction work. The agents wrote first-draft patches, yet maintainers and Microsoft’s staff still had to test and rework them: half of the 38 merged fixes were accepted on the first submission, the others after up to four rounds, and a fix usually took about 41 days from public patch to release.
Microsoft’s takeaway is that finding flaws is no longer the bottleneck. Filing 59 reports took more than four person-months of human effort, and the team warns that maintainers can drown in genuine findings, so serious bugs should come first. For people running Linux servers or Android phones, the advice stays simple: install kernel updates from your distribution or phone maker when they arrive.
