What happened
In CubeSandbox (a sandbox for running AI agents) before 0.5.1, any authenticated tenant permitted to call POST /sandboxes can supply metadata["host-mount"] with an arbitrary existing host path; the control plane performs no authorization check or allowed-prefix validation (CWE-22/CWE-284). This forces a recursive bind mount of attacker-chosen host paths into the tenant's MicroVM, granting read-write access to sensitive host files and enabling host-root compromise and cross-tenant exposure. CVSS 9.4 (Critical). Fixed in 0.5.1.
Why it matters
This is a sandbox-escape in a purpose-built multi-tenant sandbox for AI agents. A tenant owning a single sandbox can mount arbitrary host directories, read host files, run code on the host and reach co-resident tenants' data — breaking the isolation guarantee that agent-sandbox offerings are sold on. Critical severity in agent isolation infrastructure.
Attack vector
Authenticated tenant calls POST /sandboxes with a crafted host-mount metadata path pointing to an arbitrary host directory, causing a bind mount into the MicroVM
Affected systems
TencentCloud CubeSandbox before 0.5.1
Mitigation
Upgrade CubeSandbox to 0.5.1+; monitor POST /sandboxes for unusual directory paths; enforce network segmentation and egress filtering for tenant MicroVMs