Microsoft's security researchers have shown just how much AI can now do in the hunt for software flaws — and how much old code is still hiding them. A team at the company's FORGE Lab used an AI-driven system, codenamed MDASH, to scan selected parts of the Linux kernel and came back with 59 confirmed security bugs, 33 of them serious enough to earn CVE numbers.
The work, described in a Microsoft Security Blog post titled "Lessons from 59 Linux Kernel Bug Reports," was led by a single full-time researcher and one intern over three months. The scanning itself took about two weeks. The findings are a reminder that the operating system running most of the world's servers, phones and cloud infrastructure still contains flaws that have survived years of human review.
From 26,502 raw findings to 59 confirmed reports
MDASH scanned core kernel subsystems — networking, filesystems and security modules — on just two servers between late May and mid-June 2026. The first pass produced 26,502 raw findings. Automated screening agents cut that pool to 11,769 candidates, and proof-of-concept agents then tried to reproduce each one in QEMU, a virtual-machine emulator. Only 621 cases produced verifiable proof-of-concept evidence, and human experts reviewed those before 59 became public upstream reports — roughly one in ten of the reproduced cases.
The scanning was driven mainly by Anthropic's Sonnet 4.6, which handled 71.1 percent of the code-scanning work, alongside OpenAI's GPT-5.4 and GPT-5.5. Reproduction leaned heavily on GPT-5.5. Microsoft's first public report from the batch went out on June 27, 2026, and 55 of the 59 had been posted to public mailing lists by September 1.
Eight flaws hand attackers the keys
Of the 33 CVEs assigned so far, 24 are rated High or above, with CVSS scores of at least 7.0. For eight of them, the researchers confirmed working local privilege-escalation exploits in controlled laboratory conditions — flaws that would let an ordinary local user gain administrator-level control of a Linux or Android device. The affected components span IPv4 and IPv6 networking, IPsec routing, encrypted-key handling and the kernel's shared resizable hash table.
One case stands out. CVE-2026-64563 is a one-line state-reset bug in rhashtable, the kernel's shared hash-table library, which could be triggered through Netlink socket diagnostics. Researchers at NebuSec independently found the same flaw and contributed working exploits. Another, CVE-2026-74544, is a shift-out-of-bounds bug in the cls_u32 traffic classifier that had been sitting in the kernel for at least 21 years. On average, the reported bugs had been hiding for 8.1 years before Microsoft flagged them.
Discovery is now outpacing fixes
Fixes for 38 of the 59 reports have already been merged into the mainline Linux kernel, but Microsoft's researchers say the exercise exposed a new bottleneck: finding bugs is getting faster than fixing them. Converting the raw scan output into validated reports and patches took more than four person-months of expert work, and five months on, over a third of the reports are still awaiting fixes.
The team also warns that maintainers can be overwhelmed by a flood of genuine findings, and that triage should prioritise high-severity issues first. The project covered only a small fraction of the kernel, Microsoft notes, and further tuning of the approach could uncover more. For defenders, the message is blunt: AI has made vulnerability discovery cheap — the hard part is now keeping up with what it finds.



