External risk intelligence

CubeSandbox Arbitrary Host Path Access Vulnerability.

CVE advisorySeverity: CRITICAL (CVSS 9.4)

CVE-2026-120699

CubeSandbox is a platform designed for AI agents, which are commonly deployed as web-facing services or APIs. As an infrastructure component that manages tenant sandboxes, the control plane endpoint (POST /sandboxes) is typically exposed to internet-based developers or automated systems to facilitate the creation and management of these environments, making it a likely internet-facing service in common deployments.

Path Traversal

Halo Surface Signal: 4 out of 5 — likely to be public-facing.

External exposure likelihood

Horizon Alert

Summary of the vulnerability and why it matters

This vulnerability in CubeSandbox could allow an authenticated user to gain unauthorized access to host file systems, potentially leading to the execution of arbitrary code and compromise of the underlying infrastructure. The issue stems from insufficient authorization checks when handling host path requests, allowing tenants to mount arbitrary directories into their sandboxed environments.

  • Tenants can mount any host directory.
  • This allows unauthorized host access and code execution.
  • Confirm relevance and potential exposure.

Attack Path

How an attacker could exploit the issue

An attacker with authenticated access can exploit this vulnerability by sending a crafted request to the `/sandboxes` endpoint. This request allows them to specify an arbitrary host path within the `metadata["host-mount"]` field. The system then recursively binds this host path into the tenant's MicroVM, granting the attacker read-write access to host files, the ability to execute code on the host, and access to the container runtime.

  • Authenticated access to the API.
  • Supply arbitrary host path in request.
  • Host file access and code execution.

Live Threat

Current exploitation, exposure, and threat context

An authenticated user with permission to create sandboxes could manipulate the `host-mount` metadata to include arbitrary host paths. This could allow them to recursively bind mount host directories into the sandbox with read-write access, potentially leading to unauthorized access and modification of host files, execution of arbitrary code on the host, or compromise of other tenants.

  • Host file system data.
  • Arbitrary host path supplied to API.
  • Host compromise and data access.

Operational Fix

Recommended remediation, mitigation, and detection steps

In a real-world scenario, platform or infrastructure teams are likely responsible for CubeSandbox, as it manages AI agent sandboxes. The first actionable step is to identify all CubeSandbox instances, confirm their exposure and criticality, and then assign ownership to the accountable team for risk-based remediation planning.

  • Platform or Infrastructure teams own this.
  • Verify host-mount reachability and criticality.
  • Plan and coordinate remediation actions.

Supplementary metadata

Validate whether this threat affects your internet-facing exposure.

Halo Threat Intelligence helps prioritize remediation with Halo Surface Signal and H/A/L/O context. Start exposure validation with a free external attack surface trial.

Frequently asked questions

What is CubeSandbox?

CubeSandbox is a specialized infrastructure platform built to provide secure, lightweight, and concurrent execution environments specifically for AI agents. It functions by creating isolated MicroVMs where agents can operate. Its primary purpose is to allow developers or automated AI systems to manage these sandboxed workspaces dynamically through a control plane API.

What is the vulnerability in CVE-2026-120699?

This vulnerability is an improper authorization weakness (CWE-284) that allows path traversal (CWE-22) and improper privilege management (CWE-269). Because the control plane fails to validate the `host-mount` path provided by a tenant, the system mistakenly treats user-supplied paths as trusted. This allows the sandbox to mount sensitive host-level directories directly into the MicroVM with full read-write permissions, effectively breaking the isolation boundary between the sandbox and the host.

How can an attacker trigger this vulnerability?

An authenticated user triggers the bug by sending a crafted request to the POST /sandboxes API endpoint. By inserting a specific host file path into the `metadata["host-mount"]` field, the attacker forces the system to perform a recursive bind mount of that directory. Simply having access to the sandbox environment does not trigger this; the attacker must be able to influence the sandbox creation metadata, which is typically reserved for users authorized to manage their own agent environments.

How do I know if my CubeSandbox instance is at risk?

According to Halo Surface Signal, CubeSandbox is frequently deployed as a web-facing service or API to support AI agents, making it a likely internet-facing component. If your deployment allows tenants to connect to the control plane over a network, it is potentially reachable. You should evaluate your network architecture to determine if the sandbox management API is exposed to untrusted or external users, as this increases the likelihood of an attacker attempting to leverage this flaw.

What should I do to address this vulnerability?

The most effective way to secure your environment is to update CubeSandbox to version 0.5.1 or later, where the necessary authorization checks and path validation logic have been implemented. If you cannot update immediately, identify all active instances managed by your infrastructure teams, assess the reachability of your API endpoints, and restrict access to the /sandboxes creation process until the patch can be applied to prevent unauthorized directory mounting.

References