Skip to content
NewCybercrime Report 2026 · Eight Chapters · A Post Is Not An Incident → Explore The ChaptersCybercrime Report 2026 → Explore
Security NewsNews Explainer

OpenClaw Restricted-Agent Delegation: Approval Bypass Response

OpenClaw delegation approval bypass affects restricted agents under specific policies. Check the affected versions, enforce approval, and verify the fix.

By
Brandefense Research Team
Reviewed by
Brandefense Analyst Team
Published
19 November 2026
Reading time
5 min read
Layered approval boundary protecting configuration changes by a delegated AI agent

OpenClaw delegation approval bypass affects restricted agents using the delegation feature in two specified release ranges. OpenClaw says its tool could make persistent configuration changes without approval when policy denied execution and file writes. Owners should identify those configurations, move to the patched release for their branch, and review changes made before the update.

The November 19, 2026 OpenClaw advisory rates the issue High. It says a configuration change could enable host command execution, depending on local settings and lower-trust input reaching the delegation path. This is a conditional risk, not a claim that every OpenClaw deployment permits host commands.

Which OpenClaw Delegation Approval Bypass Setups Need Review?

Start with instances where restricted agents can use delegation and policy denies their execution and file writes. The advisory describes the approval bypass in that combination. It also says authenticated Gateway operators, installed plugins, and intentional local execution surfaces remain trusted within OpenClaw’s operator model.

Record each instance’s installed version, agent role, delegation setting, tool policy, channel allowlist, and Gateway users. The advisory lists versions from 2026.9.2 up to, but excluding, 2026.10.33. It separately lists versions from 2026.11.1 up to, but excluding, 2026.11.4.

Put instances that match those conditions into the vulnerability intelligence platform triage queue with their version and owner. Keep OpenClaw’s advisory attached to each decision because it defines the affected scope. That record also gives the patch owner a place to document exceptions.

Release branch Affected range in the advisory Patched destination
Extended stable 2026.9.2 through versions before 2026.10.33 2026.10.33
Stable 2026.11.1 through versions before 2026.11.4 2026.11.4

OpenClaw calls 2026.11.4 the first stable patched version and names 2026.10.33 for extended-stable users. Use the running package version, not only the intended deployment manifest, when closing the repair. If an instance falls outside the listed ranges, assess it against the advisory’s stated scope before classifying it.

Why Can a Restricted Agent Cross the Approval Boundary?

The failure is an incorrect authorization decision, identified as CWE-863 in the advisory. An agent’s normal tool policy can deny execution and file writes, yet the OpenClaw tool could still apply a persistent configuration change without the expected approval. That change can alter what the agent is permitted to do later.

The potential chain is delegation, an unapproved configuration change, and then a newly permitted host command path. The last step is conditional because the advisory ties impact to operator configuration and access by lower-trust input. Treat this as a boundary review across policy and delegation, not evidence of a completed intrusion.

The advisory assigns CVSS 3.1 score 8.8 and labels the flaw High. It does not give a CVSS 4.0 score or a CVE identifier. Neither the score nor the weakness category shows that an exposed instance was actually abused.

Brandefense Blackhole counted 456 dark-chat posts matching “OpenClaw” from August 21 through November 19, 2026. The measured unit is posts matching the product term, across that 90-day period. This count is not specific to this advisory and cannot establish exploitation or an affected deployment.

OpenClaw Delegation Approval Bypass: A Five-Step Response

Work from the instances that meet the advisory’s exact configuration conditions. Preserve the current policy and configuration before changing them, so the team can later distinguish an authorized change from an unexpected one. The following checks are our operational workflow based on OpenClaw’s stated impact and mitigations.

Restricted-agent response order
  1. Scope each GatewayRecord the running version, restricted agents, delegation availability, policy denials, and who can supply lower-trust input.
  2. Hold the boundarySet tools.exec.mode to ask, deny untrusted delegated changes, and disable the affected feature if it is unused.
  3. Patch the correct branchUpdate stable to 2026.11.4 or extended stable to 2026.10.33, then verify the running version.
  4. Review change evidenceCompare configuration history and approval records for persistent changes made by restricted agents before the fix.
  5. Restore and retestRevert unauthorized changes, confirm allowlists and policy, then test that restricted delegation still requires approval.

The first step is a scope decision: a deployment with a restricted agent but no relevant delegation path is not automatically the case described. Include shared Gateways where mutually untrusted users can reach one environment. OpenClaw advises against sharing a Gateway in that pattern as a general hardening measure.

For the second step, OpenClaw explicitly recommends tools.exec.mode set to ask before upgrading. It also recommends denying untrusted delegated changes, narrowing channel and tool allowlists, and disabling the affected feature when unnecessary. Record who approves execution while this interim control is active.

For the third step, deploy the patched version that matches the maintained branch. Verify the version from the running installation after restart and record which Gateway instances received it. A planned update or a package download alone does not prove that the process now runs the fixed build.

For the fourth step, examine the configuration history, deployment records, and any approval audit trail retained by your environment. Look for settings changed by a restricted agent despite a denied execution or file-write policy. A missing record is an investigation limit, not proof that no change happened.

For the fifth step, compare the current effective policy with the approved baseline before re-enabling delegation. Test with an approved benign change that needs the same approval boundary, and document the observed decision. The human approval and least agency checklist helps structure that review.

How Should Teams Close the Finding?

Close each affected instance only after the running version and effective controls are recorded. Keep evidence of the before and after configuration, the review of unexpected changes, and the delegated approval test. The AI agent logging guide provides a wider recordkeeping framework for this work.

For broader Gateway context, see the OpenClaw unmanaged agent gateway analysis. OpenClaw’s own advisory remains the source for this issue’s scope and patched versions. Revisit the decision when the agent policy, Gateway trust boundary, or delegation setup changes. Repeat the approval test after any change that widens a restricted agent’s tool access.

Frequently Asked Questions

Does this affect every OpenClaw agent?

No. OpenClaw scopes the advisory to restricted agents using delegation where tool policy denied execution and file writes. Check the version and local configuration.

Which OpenClaw versions fix the issue?

OpenClaw names 2026.11.4 as the first stable patched version and 2026.10.33 as the extended-stable update.

What should an owner do before the update?

OpenClaw recommends setting tools.exec.mode to ask and denying untrusted delegated changes. Narrow channel and tool allowlists and disable the feature if unused.

Sources

  1. OpenClaw security advisory GHSA-pw3x-g8v8-fvj2
Share this articleLinkedInXEmail
Brandefense logo

Stay Ahead of Digital Threats.

Take control of your digital security with an exclusive demo of our AI-driven threat intelligence platform.

Open The Full Demo Form →