Free AI Safety Workshop — in collaboration with explainx.aiSave your seat
AgentbeamBeta
AI agent vulnerability · OpenClaw

GHSA-943q-mwmv-hhvh: OpenClaw: Gateway /tools/invoke tool escalation + ACP permission auto-approval

highCVSS 3.1 8.8

Summary OpenClaw Gateway exposes an authenticated HTTP endpoint (POST /tools/invoke) intended for invoking a constrained set of tools. Two issues could combine to significantly increase blast radius in misconfigured or exposed deployments: - The HTTP gateway layer did not deny high-risk session orchestration tools by default, allowing a caller with Gateway auth to invoke tools like sessionsspawn / sessionssend and…

Published Mar 2, 2026

Affected packages

openclaw

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

Advisory details

From the GitHub Advisory Database (CC BY 4.0).

Summary

OpenClaw Gateway exposes an authenticated HTTP endpoint (POST /tools/invoke) intended for invoking a constrained set of tools. Two issues could combine to significantly increase blast radius in misconfigured or exposed deployments:

  • The HTTP gateway layer did not deny high-risk session orchestration tools by default, allowing a caller with Gateway auth to invoke tools like sessions_spawn / sessions_send and pivot into creating or controlling agent sessions.
  • ACP clients could auto-approve permission requests for risky tools with insufficient user interaction/guardrails, reducing the friction that should normally prevent silent execution or mutation.

Impact

If the Gateway is reachable by an attacker and they obtain a valid Gateway token, they may be able to:

  • Escalate from single-tool invocation to spawning/controlling sessions and reach command execution capabilities depending on tool policy and runtime environment.
  • Perform cross-session message injection via sessions_send.
  • In ACP-integrated scenarios, obtain unintended approvals for non-read/search tool permissions.

CVSS

  • CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H (8.8)

Affected versions

  • openclaw < 2026.2.14

Fixed in

  • openclaw >= 2026.2.14

Remediation

The default behavior is now hardened:

  • PR #15390: deny high-risk tools over HTTP /tools/invoke by default (with gateway.tools.{allow,deny} overrides) and harden ACP permission handling.
  • Commit bb1c3dfe1: ACP clients now prompt for any non-read/search permission request (fail closed for mutating/execution/fetch operations).
  • Commit 539689a2f: security audit warns when gateway.tools.allow re-enables default-denied HTTP tools, since this can increase RCE blast radius if the Gateway is reachable.
  • Commit 153a7644e: ACP safe-kind inference is stricter to avoid accidental auto-approval due to substring matches (still auto-approves only confident read/search).

Mitigations / deployment guidance

  • Keep the Gateway loopback-only unless you have a strong reason not to: gateway.bind="loopback" / openclaw gateway run --bind loopback.
  • Avoid exposing the Gateway directly to the public internet. Use an SSH tunnel or Tailscale to access a loopback-bound Gateway.
  • Treat opting in to default-denied HTTP tools (via gateway.tools.allow) as high-risk and audit such configurations carefully.

Credits

OpenClaw thanks @aether-ai-agent for reporting this issue and contributing remediation work.

Credited to aether-ai-agent.

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. No analysis has been written for this entry yet. Methodology

More OpenClaw vulnerabilities

All OpenClaw entries