CVE-2026-32000: OpenClaw cmd injection

CISO Take

OpenClaw versions before 2026.2.19 contain a CWE-78 OS command injection in the Lobster extension's tool execution layer: when subprocess spawning fails with EINVAL or ENOENT on Windows, the code falls back to shell: true, allowing any local user with low privileges to inject shell metacharacters in command arguments and execute arbitrary OS commands. The package ranks in the top 55% EPSS percentile with no public exploit and is absent from CISA KEV, but context elevates the risk significantly — the OpenClaw skills ecosystem was simultaneously under active abuse for malware delivery (AIID #1368), making a malicious ClawHub skill that deliberately triggers this flaw a realistic kill chain against any Windows-based AI agent deployment. Patch immediately to 2026.2.19+ (commit ba7be018); if patching is delayed, restrict Lobster tool execution to an approved command allowlist and sandbox agent processes under AppLocker or Windows Defender Application Control to deny child shell spawning.

Sources: NVD EPSS GitHub Advisory ATLAS VulnCheck

What is the risk?

CVSS 7.1 (High) with local attack vector and low privilege requirement limits direct internet-exposed risk, but AI agent workloads routinely run with elevated permissions and ingest high volumes of third-party tool inputs. The Windows-specific shell fallback path narrows scope to Windows deployments, yet the low attack complexity and zero user interaction requirement mean any malicious or compromised skill can reliably trigger exploitation. The 396 CVEs attributed to the same package signal a historically vulnerability-dense codebase, and the confirmed abuse of the OpenClaw skills ecosystem (AIID #1368) demonstrates that the third-party plugin attack surface is already being weaponized.

How does the attack unfold?

Malicious Skill Distribution

Attacker publishes a poisoned skill to ClawHub or social-engineers installation of a crafted skill package containing Lobster tool arguments with embedded shell metacharacters.

AML.T0010.005

Spawn Failure Trigger

The malicious skill invokes the Lobster extension targeting a deliberately missing or invalid binary path, inducing an EINVAL or ENOENT subprocess spawn failure on Windows.

AML.T0053

Shell Fallback Exploitation

OpenClaw's Windows fallback path re-invokes the command with shell: true, passing the unsanitized argument string directly to cmd.exe, executing the injected shell metacharacters.

AML.T0050

Arbitrary Command Execution

Attacker achieves OS command execution with the OpenClaw process privileges, enabling persistence, credential dumping, or exfiltration of agent context and API keys from the host.

AML.T0072

Malicious Skill Distribution

Attacker publishes a poisoned skill to ClawHub or social-engineers installation of a crafted skill package containing Lobster tool arguments with embedded shell metacharacters.

AML.T0010.005

Spawn Failure Trigger

The malicious skill invokes the Lobster extension targeting a deliberately missing or invalid binary path, inducing an EINVAL or ENOENT subprocess spawn failure on Windows.

AML.T0053

Shell Fallback Exploitation

OpenClaw's Windows fallback path re-invokes the command with shell: true, passing the unsanitized argument string directly to cmd.exe, executing the injected shell metacharacters.

AML.T0050

Arbitrary Command Execution

Attacker achieves OS command execution with the OpenClaw process privileges, enabling persistence, credential dumping, or exfiltration of agent context and API keys from the host.

AML.T0072

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
OpenClaw	pip	—	No patch
4 dependents 37% patched ~3d to patch Full package profile →

Do you use OpenClaw? You're affected.

How severe is it?

CVSS 3.1

7.1 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 46% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Moderate

What is the attack surface?

AV Local

AC Low

PR Low

UI None

S Unchanged

C None

I High

A High

What should I do?

6 steps

Patch: Upgrade OpenClaw to 2026.2.19 or later — commit ba7be018 contains the fix; verify the patched binary is deployed before re-enabling Lobster tool execution.
Workaround: Disable the Lobster extension or restrict it to a pre-approved command allowlist via configuration until patching is complete.
Skill hygiene: Audit installed skill inventory; remove or block unverified third-party skills from ClawHub; enforce publisher allowlisting for new skill installs.
OS-level sandbox: Run OpenClaw agent processes under a low-privilege service account; apply AppLocker or Windows Defender Application Control rules to block cmd.exe and powershell.exe as child processes of the agent binary.
Detection: Monitor EDR/SIEM for cmd.exe or powershell.exe spawned as children of the OpenClaw process, especially with concatenated or metacharacter-heavy command lines.
Review the patch diff at commit ba7be018 to understand exact argument sanitization applied and validate no parallel code paths share the same pattern.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Code Execution Supply Chain Agent Plugin AML.T0010.005 - AI Agent Tool AML.T0011.002 - Poisoned AI Agent Tool AML.T0050 - Command and Scripting Interpreter AML.T0053 - AI Agent Tool Invocation AML.T0072 - Reverse Shell

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 9 - Risk management system

ISO 42001

6.1 - Actions to address risks and opportunities 8.4 - AI system lifecycle — security testing

NIST AI RMF

MS-2.5 - Measure — AI risk identification and tracking

OWASP LLM Top 10

LLM06 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-32000?

OpenClaw versions before 2026.2.19 contain a CWE-78 OS command injection in the Lobster extension's tool execution layer: when subprocess spawning fails with EINVAL or ENOENT on Windows, the code falls back to shell: true, allowing any local user with low privileges to inject shell metacharacters in command arguments and execute arbitrary OS commands. The package ranks in the top 55% EPSS percentile with no public exploit and is absent from CISA KEV, but context elevates the risk significantly — the OpenClaw skills ecosystem was simultaneously under active abuse for malware delivery (AIID #1368), making a malicious ClawHub skill that deliberately triggers this flaw a realistic kill chain against any Windows-based AI agent deployment. Patch immediately to 2026.2.19+ (commit ba7be018); if patching is delayed, restrict Lobster tool execution to an approved command allowlist and sandbox agent processes under AppLocker or Windows Defender Application Control to deny child shell spawning.

Is CVE-2026-32000 actively exploited?

No confirmed active exploitation of CVE-2026-32000 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-32000?

1. Patch: Upgrade OpenClaw to 2026.2.19 or later — commit ba7be018 contains the fix; verify the patched binary is deployed before re-enabling Lobster tool execution. 2. Workaround: Disable the Lobster extension or restrict it to a pre-approved command allowlist via configuration until patching is complete. 3. Skill hygiene: Audit installed skill inventory; remove or block unverified third-party skills from ClawHub; enforce publisher allowlisting for new skill installs. 4. OS-level sandbox: Run OpenClaw agent processes under a low-privilege service account; apply AppLocker or Windows Defender Application Control rules to block cmd.exe and powershell.exe as child processes of the agent binary. 5. Detection: Monitor EDR/SIEM for cmd.exe or powershell.exe spawned as children of the OpenClaw process, especially with concatenated or metacharacter-heavy command lines. 6. Review the patch diff at commit ba7be018 to understand exact argument sanitization applied and validate no parallel code paths share the same pattern.

What systems are affected by CVE-2026-32000?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, AI tool execution pipelines, Windows-based AI deployments.

What is the CVSS score for CVE-2026-32000?

CVE-2026-32000 has a CVSS v3.1 base score of 7.1 (HIGH). The EPSS exploitation probability is 0.62%.

What is the AI security impact?

Affected AI Architectures

agent frameworksAI tool execution pipelinesWindows-based AI deployments

MITRE ATLAS Techniques

AML.T0010.005 AI Agent Tool

AML.T0011.002 Poisoned AI Agent Tool

AML.T0050 Command and Scripting Interpreter

AML.T0053 AI Agent Tool Invocation

AML.T0072 Reverse Shell

Compliance Controls Affected

EU AI Act: Article 9

ISO 42001: 6.1, 8.4

NIST AI RMF: MS-2.5

OWASP LLM Top 10: LLM06

What are the technical details?

Original Advisory

OpenClaw versions prior to 2026.2.19 contain a command injection vulnerability in the Lobster extension tool execution that uses Windows shell fallback with shell: true after spawn failures. Attackers can inject shell metacharacters in command arguments to execute arbitrary commands when subprocess launch fails with EINVAL or ENOENT errors.

Exploitation Scenario

An attacker publishes a malicious skill to ClawHub — the OpenClaw third-party skills marketplace implicated in AIID #1368 — or social-engineers a developer into installing it directly. The skill invokes the Lobster extension targeting a deliberately missing or invalid binary path, inducing an EINVAL or ENOENT subprocess spawn failure. OpenClaw's Windows fallback logic re-invokes the command with shell: true, passing the unsanitized argument string directly to cmd.exe. Shell metacharacters embedded in the argument (e.g., & net user backdoor P@ssw0rd /add or | powershell -enc <payload>) execute with the privileges of the OpenClaw process. The attacker achieves arbitrary command execution without requiring elevated privileges, enabling persistence via scheduled tasks, credential dumping, or exfiltration of agent context, configuration files, and API keys accessible to the process.

Weaknesses (CWE)

CWE-78 Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection')

CWE-78 — Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection'): The product constructs all or part of an OS command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended OS command when it is sent to a downstream component.

[Architecture and Design] If at all possible, use library calls rather than external processes to recreate the desired functionality.
[Architecture and Design, Operation] Run the code in a "jail" or similar sandbox environment that enforces strict boundaries between the process and the operating system. This may effectively restrict which files can be accessed in a particular directory or which commands can be executed by the software. OS-level examples include the Unix chroot jail, AppArmor, and SELinux. In general, managed code may provide some protection. For example, java.io.FilePermission in the Java SecurityManager allows the software to specify restrictions on file operations. This may not be a feasible solution, and it only limits the impact to the operating system; the rest of the application may still be subject to compromise. Be careful to avoid CWE-243 and other weaknesses related to jails.

Source: MITRE CWE corpus.