Threat reportVulnerabilityTL-2026-6577
Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts
CVE-2026-99312 is a stale-MMU-context bug in KVM's nested SVM (AMD) support: an INIT emulated while an L2 guest runs with paging disabled skips the MMU reset, and the freed nested page-table root can be forged by a guest with root to read and write host physical memory.
Key facts
| Identifiers | TL-2026-6577, CVE-2026-99312 |
|---|---|
| Also known as | KVM nested SVM stale MMU context, KVM guest-to-host escape (AMD) |
| Severity | HIGH |
| CVSS | 8.8 (CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H) |
| First disclosed | |
| EPSS | CVE-2026-99312: 0.1% (4.4th percentile) |
| Affected products | Linux Linux kernel (KVM x86, nested SVM on AMD); Red Hat Red Hat Enterprise Linux 9 and 10 kernel; Ubuntu / Canonical linux kernel (and linux-aws, linux-azure, linux-gcp, linux-kvm, linux-hwe); Debian linux kernel |
| Weaknesses (CWE) | CWE-909 |
| MITRE ATT&CK techniques | T1005 Data from Local System, T1078.004 Cloud Accounts, T1082 System Information Discovery, T1499.004 Application or System Exploitation, T1565.001 Stored Data Manipulation, T1611 Escape to Host |
| Motivation | UNKNOWN |
| Target sectors | cloud-hosting, technology, service-providers |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 13 cited sources |
Detection coverage
As of 2026-11-20, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-6577 across Splunk SPL, Microsoft KQL and Sigma, covering 13 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: KVM AMD Host Kernel Fault in Nested SVM / MMU Path (CVE-2026-99312) - Post-Exploitation Host Impact [Splunk SPL]
Detects host kernel log events (oops, BUG, general protection fault, panic, unhandled vmexit, emulation failure) that reference kvm_amd, nested SVM or KVM MMU code. Exploitation of the CVE-2026-99312 stale nested-page-table root (forged NPT entries giving host memory read/write) can corrupt host pages or crash the hypervisor, so these faults on an AMD KVM host are a high-value signal of host memory read/write, stored-data manipulation, or system-exploitation denial of service.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1611, T1068, T1005, T1565.001, T1499.004
- Data sources: linux_syslog, kernel_ring_buffer, Syslog
- Severity: high
- Confidence: medium
Splunk SPL
index=linux (sourcetype=syslog OR sourcetype=linux_messages_syslog)
("kvm_amd" OR "nested_svm" OR "kvm_mmu" OR "svm_")
("BUG:" OR "Oops" OR "general protection fault" OR "Kernel panic" OR "unhandled vmexit" OR "emulation failure")
| eval fault=case(like(_raw,"%Kernel panic%"),"panic", like(_raw,"%general protection fault%"),"gpf", like(_raw,"%BUG:%"),"bug", like(_raw,"%Oops%"),"oops", true(),"kvm_vmexit_or_emulation")
| stats count min(_time) as first_seen max(_time) as last_seen values(fault) as fault by host
| convert ctime(first_seen) ctime(last_seen)
| table host fault count first_seen last_seenMicrosoft KQL
Syslog
| where TimeGenerated > ago(24h)
| where Facility == "kern"
| where SyslogMessage has_any ("kvm_amd", "nested_svm", "kvm_mmu", "svm_")
| where SyslogMessage has_any ("BUG:", "Oops", "general protection fault", "Kernel panic", "unhandled vmexit", "emulation failure")
| summarize Events = count(), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated), Sample = any(SyslogMessage) by Computer, bin(TimeGenerated, 15m)
| project Computer, TimeGenerated, Events, FirstSeen, LastSeen, SampleSigma
title: KVM AMD Host Kernel Fault in Nested SVM or MMU Path (CVE-2026-99312)
id: 3f6a1c52-6577-4d0e-8a11-0a1b2c3d4e02
status: experimental
description: Detects host kernel faults referencing kvm_amd, nested SVM or KVM MMU code that may indicate CVE-2026-99312 exploitation or impact.
references:
- https://lists.openwall.net/linux-cve-announce/2026/11/18/2
- https://www.terra.security/blog/kvm-guest-to-host-escape
- https://www.digitalocean.com/blog/cve-2026-99312-kvm-patch
- https://nvd.nist.gov/vuln/detail/CVE-2026-99312
author: Threadlinqs Team
date: 2026-11-20
tags:
- attack.privilege-escalation
- attack.t1611
- attack.t1068
- attack.collection
- attack.t1005
- attack.impact
- attack.t1565.001
- attack.t1499.004
- cve.2026.99312
logsource:
product: linux
service: syslog
detection:
keywords_module:
- 'kvm_amd'
- 'nested_svm'
- 'kvm_mmu'
keywords_fault:
- 'BUG:'
- 'Oops'
- 'general protection fault'
- 'Kernel panic'
- 'unhandled vmexit'
- 'emulation failure'
condition: keywords_module and keywords_fault
falsepositives:
- Hardware or firmware faults on KVM hosts
- Kernel debug and virtualization test labs
level: high
Known false positives:
- Unrelated KVM or hardware faults (bad RAM, firmware bugs, buggy guest drivers) emit similar oops/BUG lines
- Debug or fuzzing kernels in virtualization test labs
Tuning notes: Scope to AMD hosts with kvm_amd nested=1 and kernels below 7.3.5 / 6.18.64 / 6.12.118; exclude lab hosts; correlate with a tenant guest that recently ran its own hypervisor (see 003) and with MCE/EDAC hardware errors to rule out faulty hardware.
- AMD KVM Nested SVM Exposure Probe or Enablement (CVE-2026-99312) - Exploitation Prerequisite [Sentinel KQL] · unlock with TL Creds
- KVM AMD Host Kernel Fault in Nested SVM / MMU Path (CVE-2026-99312) - Post-Exploitation Host Impact [Sentinel KQL] · unlock with TL Creds
- Root in Guest VM Launching a Nested KVM Hypervisor (CVE-2026-99312 Precondition) - Guest Nested-Hypervisor Staging [Sentinel KQL] · unlock with TL Creds
- AMD KVM Nested SVM Exposure Probe or Enablement (CVE-2026-99312) - Exploitation Prerequisite [Sigma] · unlock with TL Creds
- KVM AMD Host Kernel Fault in Nested SVM / MMU Path (CVE-2026-99312) - Post-Exploitation Host Impact [Sigma] · unlock with TL Creds
- Root in Guest VM Launching a Nested KVM Hypervisor (CVE-2026-99312 Precondition) - Guest Nested-Hypervisor Staging [Sigma] · unlock with TL Creds
- AMD KVM Nested SVM Exposure Probe or Enablement (CVE-2026-99312) - Exploitation Prerequisite [Splunk SPL] · unlock with TL Creds
- Root in Guest VM Launching a Nested KVM Hypervisor (CVE-2026-99312 Precondition) - Guest Nested-Hypervisor Staging [Splunk SPL] · unlock with TL Creds
Indicators of compromise
As of 2026-11-20, Threadlinqs Intelligence tracks 13 indicators of compromise for TL-2026-6577.
Sample indicators (3 of 13)
- File hash:
c7c288d178cb0b78d3464a616b16b1aa0b316049 - File hash:
f55a3e4e08dc004b4bab7c1e9323d7e7e96e266a - File hash:
24f644ea26c2ebb7b14c3baa2aa92d37fde3095f
Defanged, TLP:CLEAR. Full set of 13 indicators: unlock with TL Creds.
Technical analysis
Fixed upstream in kernels 7.3.5, 6.18.64 and 6.12.118; no in-the-wild exploitation or public exploit code is reported.
CVE-2026-99312 (kernel CVE title: "KVM: x86: Unconditionally reset MMU context when emulating INIT") affects the x86 KVM hypervisor on AMD hosts with nested virtualization enabled (the kvm_amd nested parameter, reported as on by default). Per the kernel CVE announcement of 2026-11-18 (arch/x86/kvm/x86.c), when INIT is emulated while the nested guest (L2) is running with CR0.PG=0, KVM skipped the MMU reset, leaving L1 running with a stale MMU context whose page role still includes guest_mode. The bug affects only nested SVM, because INIT is an unconditional intercept on Intel VMX and cannot be emulated while L2 runs. The fix forces an MMU reset on every emulated INIT, even if paging was already disabled. The flaw was introduced in 6.1 (commit ed129ec9057f) and back-ported to 5.15.81 (6425c590d0cc) and 6.0.11 (d40ef0a51167); the mainline fix is c7c288d178cb (v7.4-rc2). Terra Security states the omission originated in the 2022 fix for CVE-2022-3344, which switched a control-block reset path but missed the page-table pointer, leaving the bug in the kernel for just under four years.
Terra Security's researcher Oren Yomtov (reported through DigitalOcean's bug bounty and to the KVM maintainers on 2026-10-21) describes the exploitation mechanics: the VMCB nCR3 field (the nested-page-table root pointer) is left naming the freed nested-VM root table. When the guest touches new memory, KVM reuses the freed page as a bottom-level NPT table populated with host addresses of the guest's own pages; the CPU walks it as a top-level table, so an entry resolves to a guest-owned page (Page A). By writing into Page A, the guest forges second-level entries that can point to any host physical page, including host kernel memory and other tenants' VM memory. Demonstrated impact: on a DigitalOcean Droplet, about 127 GB of host RAM was mapped read-only and the host kernel version banner was recovered; on the researchers' own bare-metal AMD server with no other customers, 32 chosen bytes were written into a live host kernel page (confirmed by a witness kernel module). Prerequisites per the sources are an AMD CPU, nested virtualization enabled on the host, and root inside a guest that runs its own nested VM; sources differ on the exact privilege requirement (one AI-generated analysis cites CAP_SYS_ADMIN, which conflicts with the root-in-VM description).
Severity ratings conflict: the Linux kernel CNA scores 8.8 (CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H); Red Hat's preliminary rating is Moderate 5.6 (AV:L/AC:H/PR:L/UI:N/S:C/C:N/I:N/A:H, treated as denial of service) with CWE-909; SUSE is cited at 7.8; DigitalOcean treated it as critical. Intel-based KVM hosts and AMD hosts without nested virtualization are not affected, and many cloud providers were not vulnerable because they had not enabled nested virtualization on AMD. No public exploit code and no in-the-wild exploitation are reported, it is not in CISA KEV, EPSS is under 1%, and no detection guidance has been published. Updating guest operating systems does not help: the fix is in the host kernel and requires a reboot. DigitalOcean live-patched its fleet and states Droplet customers need no action. As of 2026-11-19/20 Debian lists everything except unstable (7.3.5-1) as vulnerable, Ubuntu lists 22.04 through 26.10 kernels as vulnerable (High) with no fix, and Red Hat marks RHEL 9/10 kernels as fix deferred (RHEL 8 not affected).
MITRE ATT&CK techniques by tactic:
Collection
Initial Access
Discovery
T1082 System Information Discovery
Impact
T1499.004 Application or System Exploitation; T1565.001 Stored Data Manipulation
Privilege Escalation
Remediation
Patches
- Upgrade host kernel to 7.3.5, 6.18.64, 6.12.118 or later (mainline fix c7c288d178cb in v7.4-rc2); Debian unstable 7.3.5-1 is fixed
- Install distribution kernels containing the fix when released (Ubuntu and RHEL 9/10 fixes pending as of 2026-11-20)
Immediate actions
- Check exposure on AMD KVM hosts: cat /sys/module/kvm_amd/parameters/nested (1 or Y means nested virtualization is on)
- If the host cannot be patched yet, set 'options kvm_amd nested=0' in /etc/modprobe.d/ (and remove kvm_amd.nested=1 from GRUB if present), warn tenants that nested VMs will stop working, then reboot the host
- Confirm the running kernel with uname -r and re-read the nested parameter after reboot
Workarounds
- Set the kvm_amd module parameter nested=0 and reboot the host; CPU-mode changes are not a mitigation
- Individual commit cherry-picking is listed by the kernel CNA but is not recommended or supported
Longer-term hardening
- Disable nested virtualization on AMD hypervisors unless a tenant workload requires it
- Track distro kernel advisories (Debian, Ubuntu, RHEL 9/10, AlmaLinux/Rocky 9/10) and schedule host reboots/live patching for multi-tenant AMD hosts
- Segregate untrusted tenants from sensitive workloads on hosts that must keep nested virtualization enabled
Affected and fixed versions
- Linux — Linux kernel (KVM x86, nested SVM on AMD)
Vulnerable versions: 5.15.81 and later 5.15.y; 6.0.11 and later 6.0.y; 6.1 and later 6.1.y/6.6.y without a fix; 6.12 before 6.12.118; 6.18 before 6.18.64; 7.3 before 7.3.5; 7.4 before 7.4-rc2
Fixed in: 6.12.118; 6.18.64; 7.3.5; 7.4-rc2 - Red Hat — Red Hat Enterprise Linux 9 and 10 kernel
Vulnerable versions: RHEL 9 kernel; RHEL 10 kernel
Fixed in: Fix deferred as of 2026-11-19; RHEL 8 not affected - Ubuntu / Canonical — linux kernel (and linux-aws, linux-azure, linux-gcp, linux-kvm, linux-hwe)
Vulnerable versions: 22.04 LTS; 24.04 LTS; 26.04 LTS; 26.10
Fixed in: No fixed package released as of 2026-11-20 - Debian — linux kernel
Vulnerable versions: bookworm 6.1.176-1 / 6.1.192-2; trixie 6.12.107-1 / 6.12.117-1; forky 7.3.4-1
Fixed in: sid 7.3.5-1
Timeline
- Oren Yomtov (Terra Security) reports the flaw to the Linux KVM maintainers; it had earlier been validated by DigitalOcean through its bug bounty and rated critical.
- KVM maintainers merge the mainline fix (c7c288d178cb) that unconditionally resets the MMU context when emulating INIT.
- Fix first tagged in upstream v7.4-rc2.
- Fixes ship in stable kernels 7.3.5, 6.18.64 and 6.12.118; DigitalOcean rolls out a live patch across its hypervisor fleet.
- CVE-2026-99312 published by the Linux kernel CNA and Terra Security publishes its technical analysis (127 GB host RAM read, 32-byte host kernel write).
- DigitalOcean publishes its disclosure post; the kernel CNA scores the CVE 8.8 while Red Hat's preliminary rating is Moderate 5.6; Debian and Ubuntu list kernels as vulnerable.
- The Hacker News and The Register report the flaw; Ubuntu and RHEL 9/10 fixes remain pending and no in-the-wild exploitation or public exploit code is reported.
Sources cited
- CVE-2026-99312: KVM: x86: Unconditionally reset MMU context when emulating INIT (linux-cve-announce) lists.openwall.net · Advisory · 2026-11-18
- Upstream fix commit c7c288d178cb (v7.4-rc2) git.kernel.org · Advisory · 2026-11-02
- Debian Security Tracker CVE-2026-99312 security-tracker.debian.org · Advisory · 2026-11-18
- Ubuntu CVE-2026-99312 ubuntu.com · Advisory · 2026-11-19
- NVD CVE-2026-99312 NVD (NIST) · Advisory · 2026-11-18
- OpenCVE CVE-2026-99312 app.opencve.io · Advisory · 2026-11-18
- Virtualization, Responsible Disclosure, and CVE-2026-99312 (DigitalOcean) digitalocean.com · Vendor report · 2026-11-19
- CVE-2026-99312: KVM Guest-to-Host Escape (Terra Security) terra.security · Research · 2026-11-18
- Red Hat CVE-2026-99312 access.redhat.com · Vendor report · 2026-11-18
- VirtFusion operator guide: CVE-2026-99312 docs.virtfusion.com · Vendor report · 2026-11-19
- KernelScan CVE-2026-99312 kernelscan.io · Research · 2026-11-18
- Linux KVM Flaw CVE-2026-99312 Lets Guest VMs Access Host Memory on AMD Hosts (The Hacker News) The Hacker News · News · 2026-11-20
- Flaws in VMware and KVM need attention (The Register) The Register · News · 2026-11-20
Previous and next in vulnerability
- F5 BIG-IP APM / iRules critical stack-based buffer overflow RCE flaws (CVE-2026-121805, CVE-2026-120292, CVE-2026-120285, CVE-2026-116953, CVE-2026-123637) and APM SAML integer underflow (CVE-2026-126822) Previous report
- OpenSSL Integer Overflow Leading to Heap Buffer Overflow in EVP_EncodeBlock() Base64 Encoding (CVE-2026-121481) Next report
Drafted by the Threadlinqs research pipeline from 13 sources; published ; last updated —.