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

CVE-2026-28470: OpenClaw has an exec allowlist bypass via command substitution/backticks inside double quotes

highCVSS 4.0 7.4 · 3.1 9.8GHSA-3hcm-ggvf-rch5

OpenClaw's optional exec approvals allowlist could be bypassed with command substitution, `$()` or backticks, placed inside double quotes. A command that looks allowlisted can run an extra command inside the substitution. The `openclaw` npm package up to 2026.2.1 is affected, only where the allowlist feature is on. Upgrade to 2026.2.2.

Published Feb 17, 2026 · Updated Mar 6, 2026 · Analysis Nov 1, 2026

What an attacker can do

In a POSIX shell, $(...) and backticks still run inside double quotes. The allowlist analysis treated a double-quoted string as inert text, so a command such as an allowlisted program with a quoted argument that contains a substitution passed the check. The shell then ran the substituted command first, with whatever arguments the attacker chose.

That is OS command injection and argument injection (CWE-78, CWE-88) past the control that was meant to restrict which commands the agent may run. Anything that can feed a command to OpenClaw's exec path, such as a prompt-injected model, can use it to run commands that were never approved, with the privileges of the OpenClaw process.

The advisory limits the scope: only setups that explicitly enable the optional exec approvals allowlist are affected, and default installs are not. The data holds two scores. CVSS 4.0 gives 7.4 (local) and CVSS 3.1 gives 9.8 (network). Use the higher figure when planning for a gateway that takes commands from remote channels.

Who is affected

The npm package openclaw at 2026.2.1 and earlier, with the optional exec approvals allowlist enabled. Installs without the allowlist are unaffected, since they have no allowlist to bypass.

npm ls -g openclaw
openclaw --version

Then check the exec approvals configuration for an allowlist.

openclaw

Ecosystem
npm
Affected
< 2026.2.2
Fixed in
2026.2.2
Check
npm ls openclaw

How to fix it

Upgrade to 2026.2.2.

npm install -g [email protected]

The fix, commit d1ecb460, rejects unescaped $() and backticks inside double quotes during allowlist analysis. The advisory names no workaround. Until you upgrade, avoid relying on the allowlist as a security boundary, and keep approval prompts on for commands you do not recognise.

How to tell if you were hit

Search exec logs and approval records for allowlisted commands whose arguments contain $( or backticks, especially inside double quotes. Check process trees on the host for children of the OpenClaw process that are not on the allowlist, such as curl, sh or nc. After the upgrade, commands with these constructs are rejected, so repeated rejections point to someone testing the bypass.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

Exec approvals allowlist bypass via command substitution/backticks inside double quotes.

Affected Packages / Versions

  • Package: openclaw (npm)
  • Affected: <= 2026.2.1
  • Fixed: >= 2026.2.2

Impact

Only affects setups that explicitly enable the optional exec approvals allowlist feature. Default installs are unaffected.

Fix

Reject unescaped $() and backticks inside double quotes during allowlist analysis.

Fix Commit(s)

  • d1ecb46076145deb188abcba8f0699709ea17198

Thanks @simecek for reporting.

Credited to simecek, stanislavfortaisle.

Questions

Are default OpenClaw installs affected by CVE-2026-28470?

No. The advisory says only setups that explicitly enable the optional exec approvals allowlist are affected.

What exactly bypasses the allowlist?

Command substitution with $() or backticks inside double quotes. The shell runs the substitution, but the allowlist analysis did not look for it.

Which version is fixed?

2026.2.2. The fix rejects unescaped $() and backticks inside double quotes during allowlist analysis.

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