NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · OpenClaw

CVE-2026-28466: OpenClaw Vulnerable to Remote Code Execution via Node Invoke Approval Bypass in Gateway

criticalCVSS 4.0 9.4 · 3.1 9.9GHSA-gv46-4xfq-jv58

OpenClaw's gateway let an authenticated client bypass node-host exec approvals by injecting internal control fields into `node.invoke` parameters. That gives remote command execution on any connected node host that supports `system.run`. Versions before 2026.2.14 are affected. Upgrade the `openclaw` npm package to 2026.2.14.

Published Mar 2, 2026 · Updated Mar 6, 2026 · Analysis Oct 31, 2026

What an attacker can do

The gateway forwarded the caller's params for the system.run node command without cleaning them. The node host treated params.approved === true and params.approvalDecision as proof that a human had already approved the command, and skipped the approval workflow. A caller could simply send those fields. The weaknesses are a confused deputy (CWE-441) and incorrect authorization (CWE-863).

The attacker must authenticate to the gateway first. The advisory names a leaked or shared gateway token, or a paired device token with operator.write, as ways in. After that, they can run arbitrary commands on every connected node host. Those hosts can be developer workstations, CI runners and servers, so one stolen token can become full compromise of each machine, including its secrets and source code. The CVSS vectors score this critical, with changed scope in the 3.1 vector, because the gateway and the node hosts are separate systems.

The advisory does not say the flaw is exploited in the wild.

Who is affected

Affected: the openclaw npm package before 2026.2.14, where a gateway has node hosts attached that support system.run. Exposure rises if the gateway is reachable from untrusted networks or if gateway or device tokens are shared widely.

Check the version on the gateway and on each node host:

npm ls openclaw

For a global install, use npm ls -g openclaw.

openclaw

Ecosystem
npm
Affected
< 2026.2.14
Fixed in
2026.2.14
Check
npm ls openclaw

How to fix it

Upgrade to openclaw 2026.2.14 or later:

npm install [email protected]

The fix strips untrusted approval control fields from system.run input. The gateway re-attaches approval flags only when params.runId points to a valid exec.approval.request record whose context matches. Approval IDs are bound to the requesting device identity, so another client cannot replay them. The gateway also forwards only an allowlisted set of system.run parameters.

The advisory's mitigations:

  • Keep the gateway off untrusted networks.
  • Rotate gateway credentials if a token or password may have leaked.
  • Block system.run at the gateway with gateway.nodes.denyCommands, or set node exec security to deny.

How to tell if you were hit

On node hosts, review system.run executions that have no matching approval request and decision in the gateway logs. Look for node.invoke calls from operator clients or device tokens you do not recognize, and for unexpected child processes under the node host. Check which gateway and device tokens exist and rotate any you cannot account for. Beam's record of agent shell commands on the node host can help show what ran, if its hooks are installed there.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

A remote code execution (RCE) vulnerability in the gateway-to-node invocation path allowed an authenticated gateway client to bypass node-host exec approvals by injecting internal control fields into node.invoke parameters.

Affected Component

  • Gateway method: node.invoke for node command system.run
  • Node host runner: exec approval gating for system.run

Impact

If an attacker can authenticate to a gateway (for example via a leaked/shared gateway token or a paired device token with operator.write), they could execute arbitrary commands on connected node hosts that support system.run. This can lead to full compromise of developer workstations, CI runners, and servers running the node host.

Technical Details

The gateway forwarded user-controlled params to node hosts without sanitizing internal approval fields. The node host treated params.approved === true and/or params.approvalDecision as sufficient to skip the approval workflow.

Fix

Patched in OpenClaw `2026.2.14`.

  • Commits:
  • 318379cdb8d045da0009b0051bd0e712e5c65e2d
  • a7af646fdab124a7536998db6bd6ad567d2b06b0
  • c1594627421f95b6bc4ad7c606657dc75b5ad0ce
  • 0af76f5f0e93540efbdf054895216c398692afcd
  • Gateway strips untrusted approval control fields from system.run user input.
  • Gateway only re-attaches approval flags when params.runId references a valid exec.approval.request record and the request context matches. Approval IDs are bound to the requesting device identity (stable across reconnects), preventing replay by other clients.
  • Gateway forwards only an allowlisted set of system.run parameters, preventing future control-field smuggling.

Mitigations

  • Upgrade to 2026.2.14 or later.
  • Restrict access to the gateway (do not expose it to untrusted networks/users).
  • Rotate gateway credentials if you suspect token/password exposure.
  • Disable remote command execution on nodes by blocking system.run at the gateway (gateway.nodes.denyCommands) and/or by configuring node exec security to deny.

Credits

OpenClaw thanks @222n5 for reporting this issue.

Credited to 222n5.

Questions

Is CVE-2026-28466 exploitable without credentials?

No. The attacker must authenticate to the gateway, for example with a leaked gateway token or a paired device token with operator.write. After that, no human approval is needed.

Is CVE-2026-28466 exploited in the wild?

The sources I read do not say so, and it is not in CISA's KEV catalog.

Can I mitigate without upgrading?

Partly. The advisory suggests blocking system.run with gateway.nodes.denyCommands, setting node exec security to deny, restricting gateway access and rotating credentials. Upgrading to 2026.2.14 is the fix.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More OpenClaw vulnerabilities

All OpenClaw entries