AI-assisted Linux kernel review yields 59 reports and 33 CVEs
A Microsoft FORGE Lab project used AI-assisted scanning to examine selected Linux kernel subsystems, turning 26,502 initial findings into 59 public upstream reports and 33 assigned CVEs. The work also confirmed working local root-privilege-escalation exploits for eight CVEs in controlled laboratory environments. The results illustrate both the reach of automated vulnerability discovery and the substantial human effort needed to validate findings and secure upstream fixes.
The project, conducted under the codename MDASH, scanned a subset of kernel code from late May to mid-June 2026. Researchers focused on security-sensitive areas including networking, filesystems and security modules, prioritizing components that could be validated using QEMU without specialized hardware. One full-time researcher and one intern led the effort.
From automated findings to verified reports
The initial scans produced 26,502 findings in about two weeks, using two servers. Researchers and screening agents reduced that pool to 11,769 candidates. Agents then attempted reproduction in QEMU; 621 cases produced proof-of-concept evidence for expert review. The process ultimately resulted in 59 upstream reports, or 9.5% of the PoC-backed cases.
Many cases did not lead to separate reports. In 266 cases, findings shared a root cause with another case and could be addressed together. Another 163 were not reported after review, mostly because their security impact was limited. A further 133 were already reported or fixed upstream. The remaining cases contributed to the 59 reports.
The gap between automated output and actionable vulnerability reports is significant for enterprise security teams evaluating AI-assisted research. A large count of scan results is not equivalent to a matching number of distinct, exploitable flaws. Reproduction, root-cause analysis, duplicate resolution and assessment of security impact all shaped the final outcome.
Eight CVEs had confirmed lab exploits
Of the 33 CVEs assigned so far, 24 were rated High or higher, defined in the report as a CVSS v3.1 score of at least 7.0. The eight CVEs with confirmed local root-privilege-escalation exploits involved several Linux kernel areas: IPv4 and IPv6 networking, IPsec routing, the GTP tunnel driver, the rhashtable library and encrypted-key handling. Some exploit scenarios required non-default settings or disabled mitigations.
One case, CVE-2026-64563, affected rhashtable, a shared resizable hash-table implementation used by kernel subsystems. The flaw involved a stale iterator pointer. A table walk can pause and resume, and the iterator retains state about the last entry it visited. If the table was resized and that entry freed while the walk was paused, a resume path could leave the pointer uncleared. A later walk operation could then follow a pointer into freed memory.
The bug dated to a 2018 change that altered when the iterator state was reset. Exploitation depended on a combination of events during a paused walk, including a resize and removal of the referenced entry. The research identified Netlink socket diagnostics as a reachable caller for an unprivileged local user on standard kernels. The report says this path could turn the flaw into a constrained kernel write and, with other exploitation techniques, local privilege escalation. The upstream fix clears the stale pointer before the affected path returns.
Microsoft also credited Yuan Tan and Frank Wu of NebuSec with independent contributions to analyzing the issue and developing working privilege-escalation exploits. The research states that the vulnerability also affected Android with SELinux enforcing, and that exploitation did not require special kernel configurations or user namespaces.
Upstream remediation remains a multi-stage process
Fixes for 38 of the 59 reports had been merged into the Linux mainline at the time described. Half of those fixes were accepted on the first submission; the others required one to four revisions. From the first public patch to a mainline release, the median time was 41 days, with the longest reported interval at 108 days. Work with maintainers on the remaining reports was ongoing.
The researchers reported an average age of 8.1 years for the bugs when first publicly reported. Six had been present for 15 to 19 years, and three for at least 20 years. The oldest highlighted example was a cls_u32 shift-out-of-bounds issue, CVE-2026-74544, which had existed for at least 21 years.
For enterprise operators, the findings reinforce the importance of tracking kernel updates and following the remediation guidance provided by Linux distributions and device vendors. The research does not provide a single deployment-level patch status for every CVE, so organizations should assess their own kernel builds and vendor advisories rather than infer that all reported issues are fixed in their environments. The project’s broader operational lesson is that automated discovery can expand research capacity, but expert validation and coordinated upstream maintenance remain essential to turning findings into security improvements.