Vulnerability  ·  2026-09-19

vm2 Node.js sandbox — sandboxed code can access host's real https.globalAgent (agent code-exec sandbox escape)

VulnerabilityHigh impactGlobalCVE-2026-92940
vm2's NodeVM sandbox, when configured to allow the 'https' builtin module, exposes the host's actual https.globalAgent object to sandboxed code instead of an isolated proxy, because certain method calls bypass the read-only wrapper and are forwarded directly to the live host object.
vm2 is a common building block used by some JavaScript-based AI agent and code-interpreter tools to sandbox LLM-generated or user-submitted code; a host-object leak in the sandbox undermines the isolation guarantee that agentic code-execution features depend on, potentially enabling privilege escalation or lateral movement from an agent's code-exec sandbox to the host process.
When a NodeVM is explicitly configured to allow require('https'), the builtin module loader wraps the host module in a read-only proxy, but method calls like Agent.prototype.on() are forwarded to the underlying real host object rather than a sandboxed copy, exposing the host process's live https.globalAgent to sandboxed/untrusted code and enabling further sandbox-escape primitives.
vm2 versions 3.11.3 through 3.11.6
Upgrade to vm2 3.11.8 or later (which resolves this and 12 related sandbox-escape CVEs). vm2's maintainers also recommend avoiding same-process JS sandboxing as a sole security boundary — use separate processes/containers/VMs for genuinely untrusted code such as LLM-agent-generated scripts.
GitHub Security Advisory GHSA-h85j-hv3c-qfgqNVD CVE-2026-92940
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →