CVE-2026-33581: OpenClaw sandbox bypass enables

CISO Take

CVE-2026-33581 is a path traversal vulnerability in OpenClaw's message tool that allows any low-privileged remote attacker to bypass localRoots sandbox validation by substituting mediaUrl or fileUrl alias parameters, reading arbitrary files on the host with no user interaction required. For AI agent deployments — where agents routinely have filesystem access to .env files, API keys, model weights, and training data — this translates immediately from a CVSS 6.5 'medium' into a practical credential exfiltration risk that the score undersells. While EPSS sits at the 58th percentile and no public exploit currently exists, the OpenClaw ecosystem has already demonstrated active adversarial interest (AIID #1368: malicious skills delivering AMOS credential stealers via ClawHub), and 396 prior CVEs in this same package signal systemic security debt warranting heightened scrutiny. Patch to OpenClaw 2026.3.24+ immediately; until patched, restrict network access to OpenClaw instances, audit all alias parameter handling, and rotate any credentials stored in directories accessible to the agent process.

Sources: NVD EPSS GitHub Advisory VulnCheck ATLAS AIID

What is the risk?

MEDIUM per CVSS 6.5, but practically higher in AI agent deployment contexts. The network-accessible, low-privilege attack vector with no user interaction required makes this straightforward for any authenticated user to exploit. The real risk amplifier is the AI agent deployment pattern: agents store and access credential files, model artifacts, and configuration data that a single path traversal request can harvest. With 396 prior CVEs in this package, OpenClaw carries a demonstrated history of security debt. No KEV listing or public exploit keeps urgency from being critical, but the low attack complexity and meaningful EPSS percentile (top 58%) warrant prompt remediation ahead of the next sprint cycle.

How does the attack unfold?

Initial Access

Attacker authenticates to the OpenClaw instance with a low-privilege account (trial user, leaked credential, or publicly accessible endpoint) to obtain API access to the message tool.

AML.T0049

Sandbox Bypass

Attacker crafts a message tool request substituting mediaUrl or fileUrl with a path traversal alias (e.g., ../../../../.env), exploiting the failure to validate alias parameters against localRoots before path resolution.

AML.T0053

Arbitrary File Read

Server resolves the alias to a path outside the sandbox and returns file contents — exposing credentials, API keys, model artifacts, training data, or system configuration in the response body.

AML.T0037

Credential Harvest and Lateral Movement

Attacker uses exfiltrated credentials to access connected systems including databases, third-party AI APIs, RAG knowledge bases, and cloud infrastructure, expanding the breach well beyond the initial OpenClaw foothold.

AML.T0083

Initial Access

Attacker authenticates to the OpenClaw instance with a low-privilege account (trial user, leaked credential, or publicly accessible endpoint) to obtain API access to the message tool.

AML.T0049

Sandbox Bypass

Attacker crafts a message tool request substituting mediaUrl or fileUrl with a path traversal alias (e.g., ../../../../.env), exploiting the failure to validate alias parameters against localRoots before path resolution.

AML.T0053

Arbitrary File Read

Server resolves the alias to a path outside the sandbox and returns file contents — exposing credentials, API keys, model artifacts, training data, or system configuration in the response body.

AML.T0037

Credential Harvest and Lateral Movement

Attacker uses exfiltrated credentials to access connected systems including databases, third-party AI APIs, RAG knowledge bases, and cloud infrastructure, expanding the breach well beyond the initial OpenClaw foothold.

AML.T0083

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

6.5 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 43% 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 None

A None

What should I do?

6 steps

Patch: upgrade to OpenClaw ≥ 2026.3.24 (commit 1d7cb6fc03552bbba00e7cffb3aa9741f5556416).
Immediate workaround if patching is delayed: disable or network-restrict the message tool endpoint.
Audit localRoots configuration to confirm no alias parameter (mediaUrl, fileUrl, or variants) can reference paths outside the defined sandbox.
Detect exploitation: scan web and agent access logs for requests containing mediaUrl or fileUrl parameters with path traversal patterns (../, %2e%2e%2f, encoded variants).
Rotate all credentials stored in directories accessible to the OpenClaw process.
Harden the runtime: run OpenClaw under a minimal-privilege OS account with filesystem access scoped to required directories only, using container isolation or chroot as defense-in-depth against future bypass attempts.

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?

Data Extraction Privacy Violation Auth Bypass Agent Plugin Framework 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

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 9 - Risk management system

ISO 42001

A.9.1 - Information security for AI systems

NIST AI RMF

MEASURE 2.5 - AI system risks and impacts are measured

OWASP LLM Top 10

LLM02:2025 - Sensitive Information Disclosure LLM06:2025 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-33581?

CVE-2026-33581 is a path traversal vulnerability in OpenClaw's message tool that allows any low-privileged remote attacker to bypass localRoots sandbox validation by substituting mediaUrl or fileUrl alias parameters, reading arbitrary files on the host with no user interaction required. For AI agent deployments — where agents routinely have filesystem access to .env files, API keys, model weights, and training data — this translates immediately from a CVSS 6.5 'medium' into a practical credential exfiltration risk that the score undersells. While EPSS sits at the 58th percentile and no public exploit currently exists, the OpenClaw ecosystem has already demonstrated active adversarial interest (AIID #1368: malicious skills delivering AMOS credential stealers via ClawHub), and 396 prior CVEs in this same package signal systemic security debt warranting heightened scrutiny. Patch to OpenClaw 2026.3.24+ immediately; until patched, restrict network access to OpenClaw instances, audit all alias parameter handling, and rotate any credentials stored in directories accessible to the agent process.

Is CVE-2026-33581 actively exploited?

No confirmed active exploitation of CVE-2026-33581 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-33581?

1. Patch: upgrade to OpenClaw ≥ 2026.3.24 (commit 1d7cb6fc03552bbba00e7cffb3aa9741f5556416). 2. Immediate workaround if patching is delayed: disable or network-restrict the message tool endpoint. 3. Audit localRoots configuration to confirm no alias parameter (mediaUrl, fileUrl, or variants) can reference paths outside the defined sandbox. 4. Detect exploitation: scan web and agent access logs for requests containing mediaUrl or fileUrl parameters with path traversal patterns (../, %2e%2e%2f, encoded variants). 5. Rotate all credentials stored in directories accessible to the OpenClaw process. 6. Harden the runtime: run OpenClaw under a minimal-privilege OS account with filesystem access scoped to required directories only, using container isolation or chroot as defense-in-depth against future bypass attempts.

What systems are affected by CVE-2026-33581?

This vulnerability affects the following AI/ML architecture patterns: AI agent frameworks, RAG pipelines, Multi-tenant AI platforms, AI-powered SaaS applications, Model serving infrastructure.

What is the CVSS score for CVE-2026-33581?

CVE-2026-33581 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.56%.

What is the AI security impact?

Affected AI Architectures

AI agent frameworksRAG pipelinesMulti-tenant AI platformsAI-powered SaaS applicationsModel serving infrastructure

MITRE ATLAS Techniques

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

Compliance Controls Affected

EU AI Act: Article 9

ISO 42001: A.9.1

NIST AI RMF: MEASURE 2.5

OWASP LLM Top 10: LLM02:2025, LLM06:2025

What are the technical details?

Original Advisory

OpenClaw before 2026.3.24 contains a sandbox bypass vulnerability in the message tool that allows attackers to read arbitrary local files by using mediaUrl and fileUrl alias parameters that bypass localRoots validation. Remote attackers can exploit this by routing file requests through unvalidated alias parameters to access files outside the intended sandbox directory.

Exploitation Scenario

An attacker with a free trial or compromised low-privilege OpenClaw account crafts a single API request to the message tool, injecting a path traversal via the fileUrl alias parameter — for example, resolving to ../../../../opt/app/.env. The localRoots validation path fails to evaluate the alias before resolution, and the server returns the environment file contents directly in the response. Within seconds the attacker extracts database credentials, Stripe webhook secrets, and third-party AI API keys. They then use database credentials to exfiltrate subscriber PII, replay Stripe webhook events to create unauthorized paid subscriptions, and invoke external AI APIs (Anthropic, OpenAI) at the victim's cost. The entire chain requires one authenticated HTTP request and leaves log artifacts indistinguishable from a legitimate tool call.

Weaknesses (CWE)

CWE-22 Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal')

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.

[Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
[Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.