Hyper-V hypervisor (hvax64.exe) double fault BSOD while Cowork VM is attached on AMD/Win11

Summary

On a Windows 11 host running Claude Code with Cowork enabled, the system has BSOD'd twice in a week with the crash occurring inside hvax64.exe (the AMD Hyper-V hypervisor), which is the same hypervisor stack that the Cowork Linux VM (claudevm.bundle) runs on. Cowork's VM was attached to Hyper-V at the time of both crashes.

Environment

| Item | Value |
| --- | --- |
| Claude Code (desktop) | 1.6608.2 |
| Cowork VM SDK | 2.1.128 |
| OS | Windows 11 Home, 10.0.26200 |
| CPU | AMD Ryzen 5 5600G (AMD-V) |
| RAM | 32 GB |
| hvax64.exe | 10.0.26100.8328 (2026/05/01) |

Crash 1 — 2026/05/09 19:55:59 JST

• Bugcheck: 0x00020001 (0x11, 0x3e58c0, 0x1007, 0xfffff83031468f00)
• WER fault bucket: NOBLOB_HYPERVISOR_ERROR_Unhandled_DoubleFault_3e58c0_IMAGE_hvax64.exe
• Minidump: C:\Windows\Minidump\050926-15437-01.dmp (3.9 MB) — available on request
• Cowork VM state at crash: VM connected (last [startVM] VM already connected log at 19:31:06; ~25 min idle before crash)
• Post-reboot: first startVM attempt timed out after 60 s, second succeeded in 52 s (slower than typical)

Crash 2 — 2026/05/05 06:58:41 JST

• Bugcheck: 0x0000007f (0x8, ...) — UNEXPECTED_KERNEL_MODE_TRAP / Double Fault
• WER fault bucket: 0x7f_8
• Minidump: C:\Windows\Minidump\050526-15312-01.dmp — available on request

Other notes

• No Microsoft-Windows-WHEA-Logger events in the 14 days surrounding either crash → no direct hardware-error evidence.
• Cowork's [KernelBugMonitor] instrumentation in cowork_vm_node.log suggests this failure mode is already on Anthropic's radar.
• Unrelated dxgkrnl LiveKernelEvent on 2026/05/04 (likely GPU TDR, separate issue).

Ask

1. Is the hvax64.exe double-fault signature a known interaction with Cowork on AMD hosts?
2. Happy to share both minidumps via whatever channel is convenient.
3. Any recommended mitigations short of disabling Cowork entirely?