CVE-2026-35638: OpenClaw priv escalation

CISO Take

OpenClaw before version 2026.3.22 contains a privilege escalation flaw in its Control UI where any session with low-level access can self-declare arbitrary privileged scopes by exploiting a device-less allow path in the trusted-proxy mechanism — no device identity verification is required. With a CVSS of 8.8 and an EPSS placing this in the top 80th percentile for exploitation likelihood, the attack surface is significant for any organization running OpenClaw-based AI agents; the low attack complexity and minimal privilege requirement (AC:L, PR:L) make exploitation straightforward once discovered. The full C:H/I:H/A:H impact triad means a successful exploit achieves complete compromise across the affected agent control plane, and the 4 downstream dependents may inherit this trust boundary flaw if they rely on OpenClaw's scope assertions. No public exploit exists and this is not in CISA KEV, but the proximity to AIID #1368 — malicious skill abuse in the OpenClaw ecosystem leading to credential exfiltration — raises the practical risk considerably. Patch to 2026.3.22 immediately using the published commits; as a compensating control, restrict Control UI network exposure and enforce external device identity verification at the perimeter.

Sources: NVD EPSS GitHub Advisory VulnCheck ATLAS AIID

What is the risk?

HIGH. CVSS 8.8 with a network-accessible, low-complexity attack requiring only low privileges to initiate. EPSS top 80th percentile indicates elevated exploitation probability relative to all published CVEs. The complete C:H/I:H/A:H impact triad means a successful exploit yields full compromise of confidentiality, integrity, and availability within the agent control plane scope. No active exploitation confirmed (not in KEV) and no public exploit or Nuclei template exists, which prevents an immediate CRITICAL rating. However, the AI agent category, the history of 396 CVEs in this package, and the direct linkage to a known real-world incident pattern (AIID #1368) sustain a HIGH operational risk classification for any enterprise running OpenClaw in an agent orchestration role.

How does the attack unfold?

Initial Access

Attacker uses low-privilege credentials to authenticate to the network-exposed OpenClaw Control UI.

AML.T0012

Exploitation

Attacker traverses the device-less allow path in the trusted-proxy mechanism, injecting self-declared privileged scopes into the session without presenting valid device identity credentials.

AML.T0049

Privilege Escalation

Session retains attacker-declared elevated scopes, bypassing device identity verification and granting administrative-level access to the agent control plane with no additional authentication challenge.

AML.T0107

Impact

Attacker modifies agent configurations, exfiltrates data processed by AI agents, invokes privileged tool calls, or pivots laterally to systems connected through the OpenClaw orchestration layer.

AML.T0086

Initial Access

Attacker uses low-privilege credentials to authenticate to the network-exposed OpenClaw Control UI.

AML.T0012

Exploitation

Attacker traverses the device-less allow path in the trusted-proxy mechanism, injecting self-declared privileged scopes into the session without presenting valid device identity credentials.

AML.T0049

Privilege Escalation

Session retains attacker-declared elevated scopes, bypassing device identity verification and granting administrative-level access to the agent control plane with no additional authentication challenge.

AML.T0107

Impact

Attacker modifies agent configurations, exfiltrates data processed by AI agents, invokes privileged tool calls, or pivots laterally to systems connected through the OpenClaw orchestration layer.

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 21% 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?

6 steps

Patch: Upgrade OpenClaw to version 2026.3.22 or later (patches at commits 630f1479c44f78484dfa21bb407cbe6f171dac87 and ccf16cd8892402022439346ae1d23352e3707e9e on GitHub).
Network segmentation: Restrict Control UI access to trusted internal networks only — block internet-facing exposure immediately as a compensating control.
Device identity enforcement: Implement external device identity verification at the proxy/gateway layer; do not rely solely on the trusted-proxy mechanism's built-in allow path.
Session audit: Review all active sessions and revoke any with scope claims that cannot be correlated to a verified device identity.
Downstream review: Audit any dependent packages or services that trust OpenClaw scope assertions and add independent verification.
Detection: Monitor access logs for sessions presenting privileged scopes without corresponding device identity records; alert on scope escalation events in the Control UI.

How is it classified?

Auth Bypass Code Execution Agent Framework AML.T0012 - Valid Accounts AML.T0049 - Exploit Public-Facing Application AML.T0091.000 - Application Access Token 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.2.2 - Roles, responsibilities and authorities for AI

NIST AI RMF

MANAGE 2.4 - Risk treatments for identified AI risks

OWASP LLM Top 10

LLM06 - Excessive Agency

Frequently Asked Questions

What is CVE-2026-35638?

OpenClaw before version 2026.3.22 contains a privilege escalation flaw in its Control UI where any session with low-level access can self-declare arbitrary privileged scopes by exploiting a device-less allow path in the trusted-proxy mechanism — no device identity verification is required. With a CVSS of 8.8 and an EPSS placing this in the top 80th percentile for exploitation likelihood, the attack surface is significant for any organization running OpenClaw-based AI agents; the low attack complexity and minimal privilege requirement (AC:L, PR:L) make exploitation straightforward once discovered. The full C:H/I:H/A:H impact triad means a successful exploit achieves complete compromise across the affected agent control plane, and the 4 downstream dependents may inherit this trust boundary flaw if they rely on OpenClaw's scope assertions. No public exploit exists and this is not in CISA KEV, but the proximity to AIID #1368 — malicious skill abuse in the OpenClaw ecosystem leading to credential exfiltration — raises the practical risk considerably. Patch to 2026.3.22 immediately using the published commits; as a compensating control, restrict Control UI network exposure and enforce external device identity verification at the perimeter.

Is CVE-2026-35638 actively exploited?

No confirmed active exploitation of CVE-2026-35638 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-35638?

1. Patch: Upgrade OpenClaw to version 2026.3.22 or later (patches at commits 630f1479c44f78484dfa21bb407cbe6f171dac87 and ccf16cd8892402022439346ae1d23352e3707e9e on GitHub). 2. Network segmentation: Restrict Control UI access to trusted internal networks only — block internet-facing exposure immediately as a compensating control. 3. Device identity enforcement: Implement external device identity verification at the proxy/gateway layer; do not rely solely on the trusted-proxy mechanism's built-in allow path. 4. Session audit: Review all active sessions and revoke any with scope claims that cannot be correlated to a verified device identity. 5. Downstream review: Audit any dependent packages or services that trust OpenClaw scope assertions and add independent verification. 6. Detection: Monitor access logs for sessions presenting privileged scopes without corresponding device identity records; alert on scope escalation events in the Control UI.

What systems are affected by CVE-2026-35638?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, AI orchestration platforms, model serving with agent control planes.

What is the CVSS score for CVE-2026-35638?

CVE-2026-35638 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 0.29%.

What is the AI security impact?

Affected AI Architectures

agent frameworksAI orchestration platformsmodel serving with agent control planes

MITRE ATLAS Techniques

AML.T0012 Valid Accounts

AML.T0049 Exploit Public-Facing Application

AML.T0091.000 Application Access Token

AML.T0107 Exploitation for Defense Evasion

Compliance Controls Affected

EU AI Act: Article 9

ISO 42001: A.2.2

NIST AI RMF: MANAGE 2.4

OWASP LLM Top 10: LLM06

What are the technical details?

Original Advisory

OpenClaw before 2026.3.22 contains a privilege escalation vulnerability in the Control UI that allows unauthenticated sessions to retain self-declared privileged scopes without device identity verification. Attackers can exploit the device-less allow path in the trusted-proxy mechanism to maintain elevated permissions by declaring arbitrary scopes, bypassing device identity requirements.

Exploitation Scenario

An adversary with low-privilege network access to an OpenClaw Control UI crafts a session request that traverses the device-less allow path in the trusted-proxy mechanism. By embedding self-declared privileged scopes in the session handshake without presenting valid device identity credentials, the attacker's session is accepted and retains administrative-level access. From this elevated position within the AI agent control plane, the adversary modifies agent configurations to redirect tool invocations to attacker-controlled endpoints, accesses sensitive data processed by agents (including credentials harvested from agent tool configurations per AIID #1368 incident pattern), and potentially deploys poisoned skills or modified agent behaviors that persist beyond the initial session — all while appearing to the authorization layer as a legitimately scoped operator session.