CVE-2026-120699: CubeSandbox AI agent sandbox let any tenant bind-mount arbitrary host paths read-write, escaping to host root (fixed 0.5.1)
TencentCloud's CubeSandbox, a MicroVM sandbox for running untrusted, LLM-generated code from AI agents, did no authorization or path allow-listing on its host-mount sandbox option before 0.5.1. Any tenant allowed to call POST /sandboxes could pass metadata["host-mount"] with any host path, including /, and Cubelet bind-mounted it as root into the tenant's MicroVM, read-write by default. That gives host file read and write, host code execution, access to the container runtime and other tenants' sandboxes, without any kernel or VMM bug.
What it is
The path flows CubeAPI to CubeMaster to Cubelet (injectHostDirMounts, prepareHostDirVolume), which runs mount --rbind as root and exposes the subtree via virtio-fs.
From inside the sandbox an attacker can read /etc/shadow, host SSH keys, Cubelet credentials, TLS keys and the CubeEgress MITM root-CA private key; read co-resident sandboxes' rootfs and memory snapshots; reach /run/containerd/containerd.sock; and write cron entries, systemd units or root's authorized_keys for host root.
This matters most when agents or end users can create sandboxes on shared nodes: the guarantee of running untrusted model-generated code safely is void.
Fixed in CubeSandbox 0.5.1 (released 11 July 2026). GitHub advisory GHSA-8frp-jfpv-5vhv; NVD published the CVE on 16 November 2026.
How to check
- Check the CubeSandbox version running on your nodes and compare with 0.5.1.
- Search CubeAPI request logs or sandbox metadata for a host-mount key, especially hostPath values like / or system directories.
How to fix
- Upgrade CubeSandbox to 0.5.1 or later.
- If untrusted tenants or agents could create sandboxes on an older version, treat those nodes as compromised: rotate host and Cubelet credentials, the CubeEgress CA key and secrets of co-resident tenants, and rebuild the nodes.