Microsoft details 59 Linux kernel bug reports from AI-assisted research

Summary
Microsoft's FORGE Lab reports 59 Linux kernel bug reports from AI-assisted scans, including 33 CVEs; 38 fixes have merged and eight CVEs had lab-confirmed local root exploits. It details how CVE-2026-64563 enables privilege escalation through a stale rhashtable pointer.
Key points
- MDASH scans generated 26,502 raw findings in about two weeks; screening and reproduction reduced these to 621 PoC-backed cases for expert review.
- The effort produced 59 public upstream reports and 33 assigned CVEs; 24 CVEs were rated High or above.
- Researchers confirmed local root privilege-escalation exploits for eight CVEs in controlled labs; some cases required non-default settings or disabled exploit mitigations.
- Fixes for 38 reports have merged into Linux mainline; reaching a mainline release took a median of 41 days from the first public patch.
- CVE-2026-64563 stems from a stale rhashtable iterator pointer after a table resize and resume; the Netlink socket-diagnostics path can turn the use-after-free into a constrained write and local privilege escalation.
- The reported bugs averaged 8.1 years old; CVE-2026-74544 had existed for at least 21 years.
- Microsoft says AI-assisted scanning and patch drafting did not replace expert validation or collaboration with Linux maintainers, which remained necessary to confirm, report, and fix findings.
Article Details
- Attack Vectors
- An unprivileged local user can pause and resume a Netlink socket diagnostic dump via recvmsg() while the rhashtable is resized and the socket at the saved iterator pointer is freed. Reallocation can make the stale pointer reference a fake socket, enabling a constrained kernel write and local privilege escalation.
- Microsoft reported confirming working local root privilege-escalation exploits for eight CVEs in controlled laboratory environments; some cases required non-default settings or disabled exploit mitigations.
- Defensive Notes
- The upstream fix for CVE-2026-64563 clears the stale iterator pointer before returning from the affected resume path.
MITRE ATT&CK
CVE
People
Frank WuThe article credits him with identifying a viable exploitation primitive and developing a working privilege-escalation exploit for CVE-2026-64563.Yuan TanThe article says he independently found CVE-2026-64563, shared its severity with Microsoft and maintainers, and later provided working Linux and Android exploits.
Products
Androida table walk, this small library flaw enabled a stable local privilege escalation on Linux—and affected Android with SELinux enforcing—without requiring special kernel configurations or user namespaces.Linux kernelAt Microsoft’s FORGE Lab, we used codename MDASH to scan a subset of Linux kernel subsystems from late May to mid-June 2026.
Tools
GPT-5.4Code scanning mainly used Sonnet 4.6 (71.1% of code-scanning tokens), GPT-5.4 (20.1%), and GPT-5.5 (7.1%).GPT-5.5Code scanning mainly used Sonnet 4.6 (71.1% of code-scanning tokens), GPT-5.4 (20.1%), and GPT-5.5 (7.1%).GPT-5.5-cyberPoC-generating agents then attempted reproduction in QEMU, mainly using GPT-5.5 (77.3% of reproduction tokens), Opus 4.6 (13.2%), and GPT-5.5-cyber (9.4%).MDASHAt Microsoft’s FORGE Lab, we used codename MDASH to scan a subset of Linux kernel subsystems from late May to mid-June 2026.Opus 4.6PoC-generating agents then attempted reproduction in QEMU, mainly using GPT-5.5 (77.3% of reproduction tokens), Opus 4.6 (13.2%), and GPT-5.5-cyber (9.4%).QEMUOur selection criteria were straightforward: we picked security-sensitive components with no obvious dependence on specialized hardware, making QEMU-based validation simple and unified.Sonnet 4.6Code scanning mainly used Sonnet 4.6 (71.1% of code-scanning tokens), GPT-5.4 (20.1%), and GPT-5.5 (7.1%).
