CVE-2026-27523: OpenClaw sandbox path traversal

CISO Take

CVE-2026-27523 is a path traversal (CWE-22) sandbox bypass in OpenClaw (versions before 2026.2.24) where an attacker with low local privileges can craft bind source paths using symlinked parent directories with non-existent leaf components to slip past the sandbox's allowed-root and blocked-path enforcement. While the CVSS score of 6.1 is medium, OpenClaw's role as an AI agent runtime makes sandbox integrity foundational — a bypass directly undermines the isolation guarantee that contains agent-driven filesystem operations. The EPSS score places this in the top 83rd percentile for exploitation likelihood, and the package's history of 396 CVEs signals an active vulnerability surface that threat actors likely monitor. No public exploit or active exploitation is confirmed, but the exploit primitive is well-understood: symlink-assisted path traversal is a classic filesystem escape requiring only moderate skill. Upgrade to 2026.2.24 immediately; if patching is blocked, restrict local user access on agent hosts and audit all bind configurations for symlinked parent directories.

Sources: NVD EPSS GitHub Advisory ATLAS VulnCheck

What is the risk?

Medium-severity (CVSS 6.1) with elevated exploitation probability (EPSS top 83rd percentile). Local access with low privileges is required, which limits the attack surface but is realistic in shared AI infrastructure or multi-tenant compute environments. The integrity impact is high (I:H) with low availability impact (A:L) and no confidentiality impact in the base CVSS vector, though real-world exploitation could expose sensitive data depending on what is accessible via escaped bind paths. The critical concern for AI teams is that sandbox isolation is a primary defense layer for AI agent runtimes — a bypass enables unauthorized writes into model loading paths, configuration directories, or host cron/systemd paths, facilitating persistence. With 396 CVEs in the same package, the overall security posture of OpenClaw warrants heightened scrutiny beyond this individual issue.

How does the attack unfold?

Initial Access

Attacker obtains a low-privilege local account on a host running OpenClaw AI agent, either via a compromised service account, shared compute access, or a malicious OpenClaw skill (as in AIID #1368).

AML.T0012

Sandbox Bypass Preparation

Attacker creates a symlink within a directory inside the allowed-root, pointing the symlink target at a sensitive directory outside the sandbox boundary (e.g., /etc or /root).

AML.T0107

Bind Validation Evasion

Attacker specifies a bind source path traversing the symlinked parent with a non-existent leaf component; OpenClaw's validation accepts the path as within allowed-root without fully resolving the symlink chain for the missing leaf.

AML.T0105

Host Filesystem Impact

Attacker creates the previously non-existent leaf component, causing the full path to resolve outside the sandbox boundary, enabling unauthorized file writes or reads on the host — achieving persistence, credential theft, or integrity compromise.

AML.T0025

Initial Access

Attacker obtains a low-privilege local account on a host running OpenClaw AI agent, either via a compromised service account, shared compute access, or a malicious OpenClaw skill (as in AIID #1368).

AML.T0012

Sandbox Bypass Preparation

Attacker creates a symlink within a directory inside the allowed-root, pointing the symlink target at a sensitive directory outside the sandbox boundary (e.g., /etc or /root).

AML.T0107

Bind Validation Evasion

Attacker specifies a bind source path traversing the symlinked parent with a non-existent leaf component; OpenClaw's validation accepts the path as within allowed-root without fully resolving the symlink chain for the missing leaf.

AML.T0105

Host Filesystem Impact

Attacker creates the previously non-existent leaf component, causing the full path to resolve outside the sandbox boundary, enabling unauthorized file writes or reads on the host — achieving persistence, credential theft, or integrity compromise.

AML.T0025

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

6.1 / 10

EPSS

0.3%

chance of exploitation in 30 days

Higher than 17% 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 Low

What should I do?

5 steps

Upgrade to OpenClaw ≥2026.2.24, which patches bind validation logic (commit b5787e4abba0dcc6baf09051099f6773c1679ec1).
If immediate patching is blocked: audit all bind source paths configured in OpenClaw and remove any paths containing symlinked parent directory components.
Apply principle of least privilege to local users on systems running OpenClaw agents — restrict who can create symlinks in directories within the allowed-root hierarchy.
Monitor filesystem activity outside expected sandbox boundaries using auditd or eBPF-based tools (e.g., Falco) with rules alerting on writes outside designated sandbox root paths.
Review the VulnCheck advisory and GitHub security advisory GHSA-m8v2-6wwh-r4gc for vendor-provided detection guidance.

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?

Auth Bypass Code Execution Agent Framework AML.T0097 - Virtualization/Sandbox Evasion AML.T0105 - Escape to Host AML.T0107 - Exploitation for Defense Evasion

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 9 - Risk management system

ISO 42001

A.6.2.4 - AI system security

NIST AI RMF

MANAGE 2.2 - Mechanisms to manage AI risks

OWASP LLM Top 10

LLM08 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-27523?

CVE-2026-27523 is a path traversal (CWE-22) sandbox bypass in OpenClaw (versions before 2026.2.24) where an attacker with low local privileges can craft bind source paths using symlinked parent directories with non-existent leaf components to slip past the sandbox's allowed-root and blocked-path enforcement. While the CVSS score of 6.1 is medium, OpenClaw's role as an AI agent runtime makes sandbox integrity foundational — a bypass directly undermines the isolation guarantee that contains agent-driven filesystem operations. The EPSS score places this in the top 83rd percentile for exploitation likelihood, and the package's history of 396 CVEs signals an active vulnerability surface that threat actors likely monitor. No public exploit or active exploitation is confirmed, but the exploit primitive is well-understood: symlink-assisted path traversal is a classic filesystem escape requiring only moderate skill. Upgrade to 2026.2.24 immediately; if patching is blocked, restrict local user access on agent hosts and audit all bind configurations for symlinked parent directories.

Is CVE-2026-27523 actively exploited?

No confirmed active exploitation of CVE-2026-27523 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-27523?

1. Upgrade to OpenClaw ≥2026.2.24, which patches bind validation logic (commit b5787e4abba0dcc6baf09051099f6773c1679ec1). 2. If immediate patching is blocked: audit all bind source paths configured in OpenClaw and remove any paths containing symlinked parent directory components. 3. Apply principle of least privilege to local users on systems running OpenClaw agents — restrict who can create symlinks in directories within the allowed-root hierarchy. 4. Monitor filesystem activity outside expected sandbox boundaries using auditd or eBPF-based tools (e.g., Falco) with rules alerting on writes outside designated sandbox root paths. 5. Review the VulnCheck advisory and GitHub security advisory GHSA-m8v2-6wwh-r4gc for vendor-provided detection guidance.

What systems are affected by CVE-2026-27523?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, sandboxed AI execution environments, multi-tenant AI compute infrastructure, AI agent deployments with filesystem bind mounts.

What is the CVSS score for CVE-2026-27523?

CVE-2026-27523 has a CVSS v3.1 base score of 6.1 (MEDIUM). The EPSS exploitation probability is 0.25%.

What is the AI security impact?

Affected AI Architectures

agent frameworkssandboxed AI execution environmentsmulti-tenant AI compute infrastructureAI agent deployments with filesystem bind mounts

MITRE ATLAS Techniques

AML.T0097 Virtualization/Sandbox Evasion

AML.T0105 Escape to Host

AML.T0107 Exploitation for Defense Evasion

Compliance Controls Affected

EU AI Act: Article 9

ISO 42001: A.6.2.4

NIST AI RMF: MANAGE 2.2

OWASP LLM Top 10: LLM08

What are the technical details?

Original Advisory

OpenClaw versions prior to 2026.2.24 contain a sandbox bind validation vulnerability allowing attackers to bypass allowed-root and blocked-path checks via symlinked parent directories with non-existent leaf paths. Attackers can craft bind source paths that appear within allowed roots but resolve outside sandbox boundaries once missing leaf components are created, weakening bind-source isolation enforcement.

Exploitation Scenario

An attacker with a low-privilege local account on a server running OpenClaw AI agents identifies that the sandbox allows bind-mounting of /data/models. They create a symlink at /home/attacker/link → /etc, then specify the bind source as /home/attacker/link/nonexistent_subdir. OpenClaw's bind validation checks that /home/attacker/link/nonexistent_subdir is beneath the allowed root but does not fully resolve the symlink chain for the non-existent leaf, so the check passes. The attacker then creates /home/attacker/link/nonexistent_subdir/cron.d (which resolves via symlink to /etc/cron.d) and writes a cron job file, achieving persistence on the host system running the AI agent. In an OpenClaw skills ecosystem context (see AIID #1368), a malicious skill could automate this sequence to establish a beachhead before exfiltrating credentials.

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.