Technical impact
CubeSandbox’s control plane fails to check authentication or validate host‑path prefixes supplied in the host‑mount metadata for the POST /sandboxes endpoint. Any authenticated caller can therefore mount any existing host directory with read/write access into a MicroVM. This grants the tenant ability to read, modify, and execute files on the host system, directly manipulate container runtime directories, and potentially compromise co‑resident tenants without invoking kernel or virtual‑machine‑monitor vulnerabilities. The weakness is a classic directory traversal combined with privilege escalation, exposing the host to arbitrary code execution.
Affected Systems
The vulnerability affects TencentCloud’s CubeSandbox installations running versions prior to 0.5.1. CubeSandbox is a lightweight, concurrent sandbox designed for AI agents, and the flaw resides in its API endpoint for sandbox creation.
Risk and Exploitability
The CVSS score of 9.4 reflects the high severity, and while a precise EPSS score is not available, the risk remains significant due to the lack of any known mitigations. The defect is not listed in the CISA KEV catalog. Exploitation requires only authenticated access to the sandbox creation API and does not depend on kernel or hypervisor bugs, making the attack vector network‑based and confined to tenants with valid credentials. The absence of public exploit code does not mitigate the intrinsic danger of the capability to mount arbitrary host directories with full privileges.