Summary
Cloud Hypervisor is vulnerable to arbitrary host file exfiltration (constrained by process privileges) when using virtio-block devices backed by raw images.
A malicious guest can overwrite its disk header with a crafted QCOW2 structure pointing to a sensitive host path. Upon the next VM boot or disk scan, the image format auto-detection parses this header and serves the host file's contents to the guest. Guest-initiated VM reboots are sufficient to trigger a disk scan and do not cause the Cloud Hypervisor process to exit. Therefore, a single VM can perform this attack without needing interaction from the management stack.
Successful exploitation requires the backing image to be either writable by the guest or sourced from an untrusted origin. Deployments utilizing only trusted, read-only images are not affected.
Critical Window
While the logic existed since 2023 (081a6eb), two recent commits (neither included in released versions) significantly increased the severity:
-
b3922db (2026-01-15): Enabled raw backing file support, allowing exfiltration of any host file (not just other qcow2 files).
-
c569a4c (2026-02-07): Removed the minimum file size constraint (64KB), allowing exfiltration of small files like SSH keys or password fragments.
Affected Revisions
| Revision Range |
Impact |
Notes |
| July 19, 2023 – Jan 15, 2026 (081a6eb to cb49595) |
High |
Host files that are valid qcow2 images can be exfiltrated (constrained by process privileges). |
| Jan 15 – Feb 10, 2026 (b3922db to a702bf1) |
Critical |
Arbitrary host file exfiltration supported (constrained by process privileges). |
| >= Feb 10, 2026 (Starting 5098322) |
Fixed |
Mitigation (backing_files=false) enabled by default. |
| >= Feb 19, 2026 (Starting a63315d) |
Hardened |
Added hardening measures: enables explicit image typing via the user interface and prevents sector-zero writes for autodetected raw images. |
Mitigation & Workarounds
1. Upgrade to Patched Versions (Recommended)
We strongly recommend upgrading to Cloud Hypervisor v51.0 or v50.1, as these releases contain both the core fix and defense-in-depth hardening measures.
2. Enable Landlock Sandboxing
If upgrading is not possible, you should use the Landlock sandboxing capability provided by Cloud Hypervisor (e.g., via the --landlock option). Landlock restricts the Cloud Hypervisor process so it can only open file paths that were explicitly provided at startup. Even if you upgrade, using Landlock is useful defense in depth against similar vulnerabilities.
3. Restrict Process Privileges & Access
As a general best practice and immediate mitigation, ensure the host environment restricts the Cloud Hypervisor process:
- Run as an unprivileged user (never
root) to restrict file access.
- Use containerization to isolate the process from the host filesystem.
- Apply strict UNIX permissions/ACLs to prevent the process from reading sensitive host files.
- Enforce read-only mounts for host resources passed to the guest when possible.
Acknowledgements
-
Reporting & Mitigation: We thank @DemiMarie for responsibly disclosing this vulnerability and for their significant engineering contributions to the fix.
-
Security Analysis: Special thanks to the Crusoe Cloud Security Team (specifically @jof) for providing deep technical analysis and validating the root cause.
-
Coordination: The security response, remediation, and disclosure process were coordinated by the Cloud Hypervisor Maintenance Team.
Summary
Cloud Hypervisor is vulnerable to arbitrary host file exfiltration (constrained by process privileges) when using virtio-block devices backed by raw images.
A malicious guest can overwrite its disk header with a crafted QCOW2 structure pointing to a sensitive host path. Upon the next VM boot or disk scan, the image format auto-detection parses this header and serves the host file's contents to the guest. Guest-initiated VM reboots are sufficient to trigger a disk scan and do not cause the Cloud Hypervisor process to exit. Therefore, a single VM can perform this attack without needing interaction from the management stack.
Successful exploitation requires the backing image to be either writable by the guest or sourced from an untrusted origin. Deployments utilizing only trusted, read-only images are not affected.
Critical Window
While the logic existed since 2023 (081a6eb), two recent commits (neither included in released versions) significantly increased the severity:
b3922db (2026-01-15): Enabled raw backing file support, allowing exfiltration of any host file (not just other qcow2 files).
c569a4c (2026-02-07): Removed the minimum file size constraint (64KB), allowing exfiltration of small files like SSH keys or password fragments.
Affected Revisions
backing_files=false) enabled by default.Mitigation & Workarounds
1. Upgrade to Patched Versions (Recommended)
We strongly recommend upgrading to Cloud Hypervisor v51.0 or v50.1, as these releases contain both the core fix and defense-in-depth hardening measures.
2. Enable Landlock Sandboxing
If upgrading is not possible, you should use the Landlock sandboxing capability provided by Cloud Hypervisor (e.g., via the
--landlockoption). Landlock restricts the Cloud Hypervisor process so it can only open file paths that were explicitly provided at startup. Even if you upgrade, using Landlock is useful defense in depth against similar vulnerabilities.3. Restrict Process Privileges & Access
As a general best practice and immediate mitigation, ensure the host environment restricts the Cloud Hypervisor process:
root) to restrict file access.Acknowledgements
Reporting & Mitigation: We thank @DemiMarie for responsibly disclosing this vulnerability and for their significant engineering contributions to the fix.
Security Analysis: Special thanks to the Crusoe Cloud Security Team (specifically @jof) for providing deep technical analysis and validating the root cause.
Coordination: The security response, remediation, and disclosure process were coordinated by the Cloud Hypervisor Maintenance Team.