CVE-2025-37957: Linux kernel · Use After Free
Source GCVE Database
Not known to be exploited · CVSS 8.8 High.
Data as of 30 Oct 2026, 16:33 UTC
Thin evidence
One scorer or none| CNA (Linux) | CVSS 3.1 | 8.8 |
|---|
CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:HExploitation evidence
Disclosed: evidence found.
CVE published 20 May 2025
Detectable: no evidence on record.
No public nuclei template indexed
Weaponized: no evidence on record.
No public exploit code indexed in Metasploit or Exploit-DB
Exploited in the wildExploited: no evidence on record.
Not in CISA KEV and no source reports exploitation
Observed by OffSeq MirageSeen by Mirage: no evidence on record.
No attack traffic recorded by OffSeq Mirage sensors
A missing step is not evidence of safety. Sources: cve.org (opens in a new tab), the Metasploit, Exploit-DB and nuclei-templates index as of 30 Oct 2026.
Description
In the Linux kernel, the following vulnerability has been resolved: KVM: SVM: Forcibly leave SMM mode on SHUTDOWN interception Previously, commit ed129ec9057f ("KVM: x86: forcibly leave nested mode on vCPU reset") addressed an issue where a triple fault occurring in nested mode could lead to use-after-free scenarios. However, the commit did not handle the analogous situation for System Management Mode (SMM). This omission results in triggering a WARN when KVM forces a vCPU INIT after SHUTDOWN interception while the vCPU is in SMM. This situation was reprodused using Syzkaller by: 1) Creating a KVM VM and vCPU 2) Sending a KVM_SMI ioctl to explicitly enter SMM 3) Executing invalid instructions causing consecutive exceptions and eventually a triple fault The issue manifests as follows: WARNING: CPU: 0 PID: 25506 at arch/x86/kvm/x86.c:12112 kvm_vcpu_reset+0x1d2/0x1530 arch/x86/kvm/x86.c:12112 Modules linked in: CPU: 0 PID: 25506 Comm: syz-executor.0 Not tainted 6.1.130-syzkaller-00157-g164fe5dde9b6 #0 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014 RIP: 0010:kvm_vcpu_reset+0x1d2/0x1530 arch/x86/kvm/x86.c:12112 Call Trace: <TASK> shutdown_interception+0x66/0xb0 arch/x86/kvm/svm/svm.c:2136 svm_invoke_exit_handler+0x110/0x530 arch/x86/kvm/svm/svm.c:3395 svm_handle_exit+0x424/0x920 arch/x86/kvm/svm/svm.c:3457 vcpu_enter_guest arch/x86/kvm/x86.c:10959 [inline] vcpu_run+0x2c43/0x5a90 arch/x86/kvm/x86.c:11062 kvm_arch_vcpu_ioctl_run+0x50f/0x1cf0 arch/x86/kvm/x86.c:11283 kvm_vcpu_ioctl+0x570/0xf00 arch/x86/kvm/../../../virt/kvm/kvm_main.c:4122 vfs_ioctl fs/ioctl.c:51 [inline] __do_sys_ioctl fs/ioctl.c:870 [inline] __se_sys_ioctl fs/ioctl.c:856 [inline] __x64_sys_ioctl+0x19a/0x210 fs/ioctl.c:856 do_syscall_x64 arch/x86/entry/common.c:51 [inline] do_syscall_64+0x35/0x80 arch/x86/entry/common.c:81 entry_SYSCALL_64_after_hwframe+0x6e/0xd8 Architecturally, INIT is blocked when the CPU is in SMM, hence KVM's WARN() in kvm_vcpu_reset() to guard against KVM bugs, e.g. to detect improper emulation of INIT. SHUTDOWN on SVM is a weird edge case where KVM needs to do _something_ sane with the VMCB, since it's technically undefined, and INIT is the least awful choice given KVM's ABI. So, double down on stuffing INIT on SHUTDOWN, and force the vCPU out of SMM to avoid any weirdness (and the WARN). Found by Linux Verification Center (linuxtesting.org) with Syzkaller. [sean: massage changelog, make it clear this isn't architectural behavior]
References
- nvd.nist.gov/vuln/detail/CVE-2025-37957 (opens in a new tab)
- git.kernel.org/stable/c/a2620f8932fa9fdabc3d78ed6efb004ca409019f (opens in a new tab)
- git.kernel.org/stable/c/d362b21fefcef7eda8f1cd78a5925735d2b3287c (opens in a new tab)
- git.kernel.org/stable/c/dc86a9cd181788f6cc4a2dbcd92ea6473dc2923f (opens in a new tab)
- git.kernel.org/stable/c/e9b28bc65fd3a56755ba503258024608292b4ab1 (opens in a new tab)
- git.kernel.org/stable/c/ec24e62a1dd3540ee696314422040180040c1e4a (opens in a new tab)