CVE-2026-32033: OpenClaw path traversal leaks files

CISO Take

OpenClaw's file system tool fails to enforce the workspaceOnly boundary when paths are prefixed with '@', allowing any authenticated low-privilege user to read arbitrary files on the host — including cloud credentials, SSH keys, and .env files. With CVSS 6.5 (Confidentiality:High, network-reachable, low complexity, no user interaction required) and an EPSS placing it in the top 75th percentile of likely-exploited CVEs, the attack is trivial to execute given any valid session. There is no confirmed active exploitation and no CISA KEV listing, but the zero-skill barrier and high-value targets (credential files, cloud config) make opportunistic exploitation realistic, particularly in multi-tenant agent platforms. Upgrade to OpenClaw 2026.2.24 or later; if patching is not immediately possible, disable file system tools in agent configuration and enforce host-level path restrictions via AppArmor or SELinux.

Sources: NVD EPSS GitHub Advisory ATLAS vulncheck.com

What is the risk?

Despite a medium CVSS of 6.5, the operational risk for AI agent deployments is elevated because the vulnerability directly neutralises the primary sandboxing control organisations rely on to isolate agent file access. The EPSS top-75 percentile indicates statistically above-average exploitation likelihood. Low attack complexity and no user-interaction requirement mean any authenticated agent session — including those obtained via credential theft, prompt injection, or compromised API keys — is sufficient to trigger exfiltration of host-level secrets.

How does the attack unfold?

Initial Access

Attacker obtains a low-privilege OpenClaw session via leaked API key, compromised multi-tenant account, or indirect prompt injection embedded in a document processed by the agent.

AML.T0012

Exploitation

Attacker invokes the file system read tool with an @-prefixed absolute path such as @/home/appuser/.aws/credentials, bypassing the workspaceOnly canonicalization check without error.

AML.T0053

Data Collection

OpenClaw resolves and returns the raw contents of the requested file outside the workspace boundary, exposing cloud credentials, SSH private keys, or application secrets.

AML.T0037

Impact

Attacker uses exfiltrated credentials for cloud account takeover, lateral movement into downstream infrastructure, or further compromise of the AI agent's operating environment.

AML.T0085

Initial Access

Attacker obtains a low-privilege OpenClaw session via leaked API key, compromised multi-tenant account, or indirect prompt injection embedded in a document processed by the agent.

AML.T0012

Exploitation

Attacker invokes the file system read tool with an @-prefixed absolute path such as @/home/appuser/.aws/credentials, bypassing the workspaceOnly canonicalization check without error.

AML.T0053

Data Collection

OpenClaw resolves and returns the raw contents of the requested file outside the workspace boundary, exposing cloud credentials, SSH private keys, or application secrets.

AML.T0037

Impact

Attacker uses exfiltrated credentials for cloud account takeover, lateral movement into downstream infrastructure, or further compromise of the AI agent's operating environment.

AML.T0085

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

0.3%

chance of exploitation in 30 days

Higher than 26% 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?

4 steps

Patch: Upgrade OpenClaw to version 2026.2.24 or later (patch commit 9ef0fc2ff8fa7b145d1e746d6eb030b1bf692260).
Workaround: If patching is not immediately possible, disable file system tools entirely in OpenClaw agent configuration, or run OpenClaw under a dedicated least-privilege OS user with file access restricted to workspace paths via AppArmor profile or SELinux policy.
Detection: Audit OpenClaw tool invocation logs for requests containing '@/' path patterns; deploy host-based file integrity monitoring alerting on reads of sensitive paths (e.g., ~/.ssh, ~/.aws, /etc/shadow) by the OpenClaw process.
Scope assessment: Inventory all OpenClaw instances using workspaceOnly=true, prioritising those co-located with cloud provider credential files or secrets management agents.

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 Auth Bypass Agent Plugin AML.T0037 - Data from Local System AML.T0049 - Exploit Public-Facing Application AML.T0053 - AI Agent Tool Invocation AML.T0085 - Data from AI Services

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art. 9 - Risk management system

ISO 42001

A.8.2 - AI system testing and validation

NIST AI RMF

MS-2.5 - AI system risks or impacts are measured or estimated

OWASP LLM Top 10

LLM06 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-32033?

OpenClaw's file system tool fails to enforce the workspaceOnly boundary when paths are prefixed with '@', allowing any authenticated low-privilege user to read arbitrary files on the host — including cloud credentials, SSH keys, and .env files. With CVSS 6.5 (Confidentiality:High, network-reachable, low complexity, no user interaction required) and an EPSS placing it in the top 75th percentile of likely-exploited CVEs, the attack is trivial to execute given any valid session. There is no confirmed active exploitation and no CISA KEV listing, but the zero-skill barrier and high-value targets (credential files, cloud config) make opportunistic exploitation realistic, particularly in multi-tenant agent platforms. Upgrade to OpenClaw 2026.2.24 or later; if patching is not immediately possible, disable file system tools in agent configuration and enforce host-level path restrictions via AppArmor or SELinux.

Is CVE-2026-32033 actively exploited?

No confirmed active exploitation of CVE-2026-32033 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-32033?

1. Patch: Upgrade OpenClaw to version 2026.2.24 or later (patch commit 9ef0fc2ff8fa7b145d1e746d6eb030b1bf692260). 2. Workaround: If patching is not immediately possible, disable file system tools entirely in OpenClaw agent configuration, or run OpenClaw under a dedicated least-privilege OS user with file access restricted to workspace paths via AppArmor profile or SELinux policy. 3. Detection: Audit OpenClaw tool invocation logs for requests containing '@/' path patterns; deploy host-based file integrity monitoring alerting on reads of sensitive paths (e.g., ~/.ssh, ~/.aws, /etc/shadow) by the OpenClaw process. 4. Scope assessment: Inventory all OpenClaw instances using workspaceOnly=true, prioritising those co-located with cloud provider credential files or secrets management agents.

What systems are affected by CVE-2026-32033?

This vulnerability affects the following AI/ML architecture patterns: AI agent frameworks, Agentic pipelines with file system access, CI/CD pipelines using AI agents, Multi-tenant AI agent platforms.

What is the CVSS score for CVE-2026-32033?

CVE-2026-32033 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.34%.

What is the AI security impact?

Affected AI Architectures

AI agent frameworksAgentic pipelines with file system accessCI/CD pipelines using AI agentsMulti-tenant AI agent platforms

MITRE ATLAS Techniques

AML.T0037 Data from Local System

AML.T0049 Exploit Public-Facing Application

AML.T0053 AI Agent Tool Invocation

AML.T0085 Data from AI Services

Compliance Controls Affected

EU AI Act: Art. 9

ISO 42001: A.8.2

NIST AI RMF: MS-2.5

OWASP LLM Top 10: LLM06

What are the technical details?

Original Advisory

OpenClaw versions prior to 2026.2.24 contain a path traversal vulnerability where @-prefixed absolute paths bypass workspace-only file-system boundary validation due to canonicalization mismatch. Attackers can exploit this by crafting @-prefixed paths like @/etc/passwd to read files outside the intended workspace boundary when tools.fs.workspaceOnly is enabled.

Exploitation Scenario

An attacker with a low-privilege OpenClaw session — obtained via a leaked API key, a compromised multi-tenant platform account, or triggered indirectly through prompt injection in ingested document content — invokes the file system read tool with @/home/appuser/.aws/credentials as the path argument. The workspaceOnly validator compares the raw @-prefixed string against the workspace root without canonicalising or stripping the '@' prefix, so the boundary check passes silently. OpenClaw reads and returns the AWS credentials file to the attacker. In an automated agentic pipeline processing user-submitted documents, a malicious document embedding instructions to read @/etc/shadow could trigger this path traversal without any direct operator action, exfiltrating credentials autonomously during a routine document analysis job.

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.