Microsoft MDASH: Filtering 26,502 AI Scan Results Down to 59 Kernel Bugs and Eight Root Exploits

1. Basic Information

• Article Name: Lessons from 59 Linux Kernel Bug Reports
• Publisher: Microsoft Security
• Publication Date: 2026-11-11
• Collection Date: 2026-11-12T07:30:45+09:00
• Original Source: Microsoft Security
• Related Sources: None
• Related CVEs: CVE-2026-64554, CVE-2026-64563, CVE-2026-64580, CVE-2026-74561, CVE-2026-74562, CVE-2026-89789, CVE-2026-98222, CVE-2026-98364
• Related Products and Technologies: Linux kernel, Android kernel, MDASH, QEMU
• Severity: High
• Severity Justification: The research demonstrates the scale and limitations of AI vulnerability discovery through specific counts and reproduction steps, leading to root exploits affecting Linux and Android as well as 38 mainline fixes.

2. Executive Summary

The Microsoft FORGE Lab narrowed down 26,502 AI scan results to 59 upstream bugs through screening, QEMU PoCs, and expert reviews, resulting in 33 CVEs and confirmed local root exploits in a lab environment for 8 vulnerabilities. Human verification, rather than AI output alone, proved to be the primary bottleneck.

3. Research Objectives

• Determine how far AI-driven kernel source scanning can be scaled into reproducible bug reports, CVEs, and patches.
• Understand how to filter large volumes of raw findings into actionable evidence for experts.

4. Proposed Methodology and Inputs

Proposed Methodology

• Use MDASH to scan the source code of Linux kernel subsystems, including networking, filesystems, and security modules, then filter the findings using confidence levels and a screening agent.
• Use a PoC agent to attempt reproduction in QEMU, followed by expert reviews to assess root causes, duplicates, and severity before submitting reports upstream.

Inputs

• Selected Linux kernel subsystems.
• Scanning via Sonnet 4.6, GPT-5.4, and GPT-5.5, with PoC generation focused on GPT-5.5.
• Two servers (64 cores / 256 GB and 32 cores / 128 GB) running a QEMU validation environment.

5. Experiment and Workflow

• Filtered 26,502 raw findings down to 11,769 screened candidates, generating 621 PoC evidence items that led to 59 upstream reports after expert review.
• Collaborated with maintainers to advance reproduction, root cause analysis, patching, and backporting.

6. Results

• Out of 59 reports, 33 received CVEs, and 38 fixes were merged into the mainline kernel.
• Local root exploits were confirmed in a controlled lab for 8 CVEs, with some requiring non-default settings or disabled mitigations.
• Most raw findings were not directly reportable; human reviewers narrowed 621 PoC evidence items down to 59 reports.
• The verification team consisted of one full-time researcher and one intern over three months, indicating that human verification is the bottleneck relative to automated scanning scale.

7. Success Conditions and Scope

• Ability to reproduce the target subsystem in QEMU.
• Expert verification of PoC evidence up to the root cause and impact.
• Coordination of patches and upstream reports with maintainers.

8. Facts, Inference, and Hypothesis

Facts

• Scans conducted between late May and mid-June 2026 generated 26,502 raw findings across two servers.
• Confidence filtering and screening agents reduced the count to 11,769 candidates, after which a PoC agent attempted QEMU reproduction and experts reviewed 621 PoC evidence items.
• This process ultimately led to 59 public upstream reports, 33 CVEs, and 38 mainline merges.
• Microsoft confirmed local root privilege-escalation exploit behavior in a controlled lab for 8 vulnerabilities, though some required non-default settings or disabled mitigations. The developers of the exploits varied by case, with some developed and provided by external researchers.
• The 59 bugs had existed for an average of 8.1 years, with the oldest dating back over 21 years.
• According to Microsoft, CVE-2026-64563 is reachable by unprivileged users in a standard kernel without special kernel settings or user namespaces, and an exploit was verified to work even on Android with SELinux enforcing (confirmed after receiving Linux and Android exploits from NebuSec researchers).

Inference

• While AI can accelerate broad candidate generation and PoC creation, expert review remains essential for determining root causes, eliminating duplicates, assessing severity, and developing upstream patches.

Hypothesis

• None

9. Practical Implications

• The value of an AI scanner should be measured by its evidence pipeline—its ability to lead to reproduction, root causes, and patches—rather than its raw finding count.
• While older kernel bugs can be rediscovered using AI and QEMU, actual risk varies based on distribution backports and configurations.

10. Limitations

• The study does not cover all kernel subsystems, hardware architectures, or distributions.
• Some of the 8 exploits require non-default settings or disabled security mitigations.
• The raw finding count does not represent independent vulnerabilities, as it includes false positives and duplicates.

11. Guidance for Application

• Do not use raw AI scan results directly to assess severity; instead, document reproduction results, root causes, affected configurations, and patches.
• Determine priorities by cross-referencing distribution advisories with running kernel configurations.
• Run PoCs only in isolated lab environments and never apply them to production systems.

12. Required Evidence

• Scanner prompts, models, versions, source commits, configurations, raw findings, and screening rationales.
• QEMU images, reproduction steps, crash or write primitives, exploit prerequisites, and mitigation states.
• Upstream reports, CVEs, fix commits, and distribution backports.

13. MITRE ATT&CK Mapping

Not applicable, as the primary purpose is not mapping to active attack campaigns or the evidence is insufficient.

14. Uncertainties and Additional Investigation

• Distribution-specific backport status for all 59 bugs, the proportion of vulnerabilities present in the scanned code that MDASH missed (its false-negative rate), and reproducibility across different models and hardware remain undetermined.

15. Implications for Defenders

When updating kernels, organizations should verify distribution backports and configurations alongside CVE numbers, treating AI scan results as part of an evidence pipeline that requires PoCs, root cause analysis, and human review.

16. Summary by Role

• For SOCs: Monitor unprivileged processes, kernel crashes, and privilege transitions associated with local exploitation of the eight vulnerabilities.
• For Administrators: Review distribution advisories and backports, update affected kernels, and restrict unnecessary namespaces and modules.
• For End Users: This research does not demonstrate that AI can automatically compromise Linux as a whole. Check with your administrators whether patched kernel versions have been deployed.