CVE-2026-33573: OpenClaw workspace boundary bypass,

CISO Take

OpenClaw's gateway agent RPC fails to validate the `spawnedBy` and `workspaceDir` parameters, allowing any authenticated operator holding `operator.write` permission to override workspace boundaries and execute arbitrary file and exec operations from any directory the process can reach—no complex setup required. With a CVSS of 8.8, a network-accessible attack vector, low privilege requirement, and EPSS in the top 79th percentile, this is exploitable by any compromised or malicious insider with operator-level access, and OpenClaw's history of 396 CVEs signals systemic security debt that elevates the risk profile. While there is no confirmed public exploit and the CVE is absent from CISA KEV, the trivial exploitation path and the parallels to real-world credential exfiltration abuse in the OpenClaw ecosystem (AIID #1368) warrant treating this urgently. Upgrade to OpenClaw 2026.3.11 or later immediately; as an interim control, restrict `operator.write` grants to the minimum necessary set and audit gateway RPC logs for `workspaceDir` values traversing outside expected workspace roots.

Sources: NVD EPSS GitHub Advisory VulnCheck ATLAS

What is the risk?

HIGH. CVSS 8.8 with a network-accessible attack vector and low privilege requirement means any authenticated operator can exploit this without additional complexity. The EPSS top-79th-percentile score indicates elevated exploitation probability relative to the broader CVE landscape even without a confirmed public exploit. Organizations deploying OpenClaw in shared or multi-tenant agent environments face the greatest exposure, as a single compromised or malicious operator account becomes a full filesystem pivot. The 396 prior CVEs in this package indicate a pattern of poor security hygiene that should factor into dependency risk decisions independent of this specific CVE.

How does the attack unfold?

Initial Access

Attacker authenticates to the OpenClaw gateway using a compromised or malicious operator account that holds the operator.write permission.

AML.T0012

Exploitation

Attacker crafts an RPC call to the gateway agent supplying attacker-controlled spawnedBy and workspaceDir values (e.g., workspaceDir set to '/') to trigger the authorization bypass.

AML.T0049

Workspace Escape

OpenClaw processes the tampered RPC without validating workspace boundaries, granting the attacker file read, write, and exec access to any directory reachable by the OpenClaw process user.

AML.T0105

Impact

Attacker exfiltrates LLM API keys, RAG database credentials, agent memory, and model configuration files, or executes arbitrary commands to establish persistence or pivot to connected infrastructure.

AML.T0086

Initial Access

Attacker authenticates to the OpenClaw gateway using a compromised or malicious operator account that holds the operator.write permission.

AML.T0012

Exploitation

Attacker crafts an RPC call to the gateway agent supplying attacker-controlled spawnedBy and workspaceDir values (e.g., workspaceDir set to '/') to trigger the authorization bypass.

AML.T0049

Workspace Escape

OpenClaw processes the tampered RPC without validating workspace boundaries, granting the attacker file read, write, and exec access to any directory reachable by the OpenClaw process user.

AML.T0105

Impact

Attacker exfiltrates LLM API keys, RAG database credentials, agent memory, and model configuration files, or executes arbitrary commands to establish persistence or pivot to connected infrastructure.

AML.T0086

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

8.8 / 10

EPSS

0.3%

chance of exploitation in 30 days

Higher than 22% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What is the attack surface?

AV Network

AC Low

PR Low

UI None

S Unchanged

C High

I High

A High

What should I do?

1 step

1) Upgrade to OpenClaw 2026.3.11 or later immediately—this release patches the RPC parameter validation logic. 2) As an interim control, restrict operator.write permission to only fully trusted, human-reviewed accounts; remove it from service accounts or automation identities where not strictly necessary. 3) Apply OS-level filesystem sandboxing (Linux namespaces, seccomp, or chroot jails) to constrain what the OpenClaw process can access regardless of workspace configuration. 4) Audit historical RPC logs for calls containing workspaceDir values with path traversal patterns (../, absolute paths outside expected workspace roots, or /). 5) Rotate any secrets or API keys stored in directories accessible to the OpenClaw process user. 6) Assess whether the four downstream dependents in your environment also expose this RPC surface and apply equivalent controls.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact total

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

How is it classified?

Auth Bypass Code Execution Data Extraction Agent Framework AML.T0012 - Valid Accounts AML.T0037 - Data from Local System AML.T0049 - Exploit Public-Facing Application AML.T0053 - AI Agent Tool Invocation AML.T0083 - Credentials from AI Agent Configuration AML.T0086 - Exfiltration via AI Agent Tool Invocation AML.T0105 - Escape to Host

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 9 - Risk management system

ISO 42001

6.1.2 - AI risk assessment 8.4 - AI system operation

NIST AI RMF

GOVERN 1.2 - Policies and procedures for AI risk MANAGE 2.2 - Mechanisms to respond to AI risks

OWASP LLM Top 10

LLM08:2025 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-33573?

OpenClaw's gateway agent RPC fails to validate the `spawnedBy` and `workspaceDir` parameters, allowing any authenticated operator holding `operator.write` permission to override workspace boundaries and execute arbitrary file and exec operations from any directory the process can reach—no complex setup required. With a CVSS of 8.8, a network-accessible attack vector, low privilege requirement, and EPSS in the top 79th percentile, this is exploitable by any compromised or malicious insider with operator-level access, and OpenClaw's history of 396 CVEs signals systemic security debt that elevates the risk profile. While there is no confirmed public exploit and the CVE is absent from CISA KEV, the trivial exploitation path and the parallels to real-world credential exfiltration abuse in the OpenClaw ecosystem (AIID #1368) warrant treating this urgently. Upgrade to OpenClaw 2026.3.11 or later immediately; as an interim control, restrict `operator.write` grants to the minimum necessary set and audit gateway RPC logs for `workspaceDir` values traversing outside expected workspace roots.

Is CVE-2026-33573 actively exploited?

No confirmed active exploitation of CVE-2026-33573 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-33573?

1) Upgrade to OpenClaw 2026.3.11 or later immediately—this release patches the RPC parameter validation logic. 2) As an interim control, restrict `operator.write` permission to only fully trusted, human-reviewed accounts; remove it from service accounts or automation identities where not strictly necessary. 3) Apply OS-level filesystem sandboxing (Linux namespaces, seccomp, or chroot jails) to constrain what the OpenClaw process can access regardless of workspace configuration. 4) Audit historical RPC logs for calls containing `workspaceDir` values with path traversal patterns (`../`, absolute paths outside expected workspace roots, or `/`). 5) Rotate any secrets or API keys stored in directories accessible to the OpenClaw process user. 6) Assess whether the four downstream dependents in your environment also expose this RPC surface and apply equivalent controls.

What systems are affected by CVE-2026-33573?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, AI automation pipelines, multi-tenant agent platforms, model serving infrastructure, CI/CD with AI agents.

What is the CVSS score for CVE-2026-33573?

CVE-2026-33573 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 0.30%.

What is the AI security impact?

Affected AI Architectures

agent frameworksAI automation pipelinesmulti-tenant agent platformsmodel serving infrastructureCI/CD with AI agents

MITRE ATLAS Techniques

AML.T0012 Valid Accounts

AML.T0037 Data from Local System

AML.T0049 Exploit Public-Facing Application

AML.T0053 AI Agent Tool Invocation

AML.T0083 Credentials from AI Agent Configuration

AML.T0086 Exfiltration via AI Agent Tool Invocation

AML.T0105 Escape to Host

Compliance Controls Affected

EU AI Act: Article 9

ISO 42001: 6.1.2, 8.4

NIST AI RMF: GOVERN 1.2, MANAGE 2.2

OWASP LLM Top 10: LLM08:2025

What are the technical details?

Original Advisory

OpenClaw before 2026.3.11 contains an authorization bypass vulnerability in the gateway agent RPC that allows authenticated operators with operator.write permission to override workspace boundaries by supplying attacker-controlled spawnedBy and workspaceDir values. Remote operators can escape the configured workspace boundary and execute arbitrary file and exec operations from any process-accessible directory.

Exploitation Scenario

An attacker who has compromised an operator account—via phishing, credential stuffing, or an insider threat—with `operator.write` permission sends a crafted RPC call to OpenClaw's gateway agent, setting `workspaceDir` to `/` or another sensitive path and `spawnedBy` to a legitimate-looking agent identity. OpenClaw processes the request without validating workspace boundaries, granting the attacker read, write, and exec access to the entire filesystem reachable by the OpenClaw process. In an AI agent environment, the attacker can then exfiltrate model configuration files, API keys for LLM providers, RAG database credentials, and agent memory—then pivot to connected infrastructure or poison agent tools for persistence. This attack surface directly mirrors the OpenClaw skills ecosystem abuse documented in AIID #1368, where the same agent platform was leveraged to deliver credential-exfiltrating malware.

Weaknesses (CWE)

CWE-668 Exposure of Resource to Wrong Sphere

CWE-668 — Exposure of Resource to Wrong Sphere: The product exposes a resource to the wrong control sphere, providing unintended actors with inappropriate access to the resource.

Source: MITRE CWE corpus.