GHSA-jhpv-5j76-m56h: OpenClaw auth bypass leaks

CISO Take

OpenClaw 2026.4.9 contains a policy enforcement mismatch where sender and group-scoped `read` denials are not propagated to the host-media attachment read helper, allowing any channel participant explicitly denied `read` access to still retrieve files from the host filesystem via the outbound media path. In multi-tenant or mixed-trust channel deployments — a common pattern in AI agent orchestration — this silently nullifies fine-grained access controls without any error surfaced to administrators. With only 4 downstream dependents and no public exploit or CISA KEV entry, mass exploitation is unlikely, but insider-threat and untrusted-partner scenarios are directly enabled in any deployment combining global host-read permission with sender-scoped denials. Patch to OpenClaw 2026.4.10 immediately; if patching is not feasible, disable host read and filesystem root expansion at the global and agent level as a temporary workaround.

Sources: GitHub Advisory ATLAS

What is the risk?

Medium risk overall, but high risk in specific configurations. The vulnerability requires the attacker to already be an authenticated channel participant, limiting the attack surface to insiders or compromised accounts. However, in deployments that specifically rely on sender-scoped policy to segment trust levels — precisely the architecture this feature supports — the bypass completely undermines the intended security control. The CWE-863 class of authorization bypass in agent tool policy enforcement is particularly concerning given the industry trend toward multi-agent orchestration with mixed-trust participant pools.

How does the attack unfold?

Initial Position

Attacker is an authenticated but low-trust channel participant whose `read` access has been explicitly denied via `toolsBySender` or group policy in the target OpenClaw deployment.

AML.T0012

Policy Bypass Trigger

Attacker sends a crafted message triggering the host-media attachment loading path, which installs the host `readFile` capability without checking the sender's policy context.

AML.T0053

Unauthorized File Access

The host-media read helper executes with the OpenClaw process's full filesystem permissions, loading local files (credentials, config, env vars) without sender authorization checks.

AML.T0037

Data Exfiltration

Sensitive host files are returned to the attacker through the normal outbound media response channel, indistinguishable from a legitimate media attachment.

AML.T0086

Initial Position

Attacker is an authenticated but low-trust channel participant whose `read` access has been explicitly denied via `toolsBySender` or group policy in the target OpenClaw deployment.

AML.T0012

Policy Bypass Trigger

Attacker sends a crafted message triggering the host-media attachment loading path, which installs the host `readFile` capability without checking the sender's policy context.

AML.T0053

Unauthorized File Access

The host-media read helper executes with the OpenClaw process's full filesystem permissions, loading local files (credentials, config, env vars) without sender authorization checks.

AML.T0037

Data Exfiltration

Sensitive host files are returned to the attacker through the normal outbound media response channel, indistinguishable from a legitimate media attachment.

AML.T0086

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
OpenClaw	npm	>= 2026.4.9, < 2026.4.10	`2026.4.10`
4 dependents 36% patched ~3d to patch Full package profile →

Do you use OpenClaw? You're affected.

How severe is it?

CVSS 3.1

N/A

EPSS

N/A

Exploitation Status

No known exploitation

Sophistication

Moderate

What should I do?

5 steps

Patch: Upgrade openclaw to >= 2026.4.10 immediately. The fix (PR #64459, commit c949af9) threads sender, session, channel, and account context into outbound media access resolution and intersects host-media read capability creation with existing group tool policy.
Workaround (if immediate patching is not possible): Disable host read and filesystem root expansion at the global and agent level to remove the vulnerable capability entirely.
Audit: Enumerate all deployments using toolsBySender or group policy to deny read for any participant — these are the specifically affected configurations.
Detection: Review outbound media logs for file reads from sensitive paths (e.g., ~/.env, credential stores, /etc) attributed to low-trust or sender-restricted participants.
For defense-in-depth, additional attachment canonicalization hardening shipped in 2026.4.14 — consider targeting that version.

How is it classified?

Auth Bypass Data Extraction Agent Plugin AML.T0037 - Data from Local System AML.T0053 - AI Agent Tool Invocation AML.T0085.001 - AI Agent Tools AML.T0107 - Exploitation for Defense Evasion

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art. 9 - Risk management system

ISO 42001

A.6.1.3 - Access control for AI systems

NIST AI RMF

GOVERN-1.1 - Policies and procedures are in place to address AI risks

OWASP LLM Top 10

LLM07 - Insecure Plugin Design LLM08 - Excessive Agency

Frequently Asked Questions

What is GHSA-jhpv-5j76-m56h?

OpenClaw 2026.4.9 contains a policy enforcement mismatch where sender and group-scoped `read` denials are not propagated to the host-media attachment read helper, allowing any channel participant explicitly denied `read` access to still retrieve files from the host filesystem via the outbound media path. In multi-tenant or mixed-trust channel deployments — a common pattern in AI agent orchestration — this silently nullifies fine-grained access controls without any error surfaced to administrators. With only 4 downstream dependents and no public exploit or CISA KEV entry, mass exploitation is unlikely, but insider-threat and untrusted-partner scenarios are directly enabled in any deployment combining global host-read permission with sender-scoped denials. Patch to OpenClaw 2026.4.10 immediately; if patching is not feasible, disable host read and filesystem root expansion at the global and agent level as a temporary workaround.

Is GHSA-jhpv-5j76-m56h actively exploited?

No confirmed active exploitation of GHSA-jhpv-5j76-m56h has been reported, but organizations should still patch proactively.

How to fix GHSA-jhpv-5j76-m56h?

1. Patch: Upgrade openclaw to >= 2026.4.10 immediately. The fix (PR #64459, commit c949af9) threads sender, session, channel, and account context into outbound media access resolution and intersects host-media read capability creation with existing group tool policy. 2. Workaround (if immediate patching is not possible): Disable host read and filesystem root expansion at the global and agent level to remove the vulnerable capability entirely. 3. Audit: Enumerate all deployments using `toolsBySender` or group policy to deny `read` for any participant — these are the specifically affected configurations. 4. Detection: Review outbound media logs for file reads from sensitive paths (e.g., ~/.env, credential stores, /etc) attributed to low-trust or sender-restricted participants. 5. For defense-in-depth, additional attachment canonicalization hardening shipped in 2026.4.14 — consider targeting that version.

What systems are affected by GHSA-jhpv-5j76-m56h?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, multi-tenant AI deployments, AI agent orchestration.

What is the CVSS score for GHSA-jhpv-5j76-m56h?

No CVSS score has been assigned yet.

What is the AI security impact?

Affected AI Architectures

agent frameworksmulti-tenant AI deploymentsAI agent orchestration

MITRE ATLAS Techniques

AML.T0037 Data from Local System

AML.T0053 AI Agent Tool Invocation

AML.T0085.001 AI Agent Tools

AML.T0107 Exploitation for Defense Evasion

Compliance Controls Affected

EU AI Act: Art. 9

ISO 42001: A.6.1.3

NIST AI RMF: GOVERN-1.1

OWASP LLM Top 10: LLM07, LLM08

What are the technical details?

Original Advisory

## Summary OpenClaw's outbound host-media attachment read helper could enable host-local file reads based on global or agent-level read access without also honoring sender and group-scoped tool policy. In channel deployments that used `toolsBySender` or group policy to deny `read` for less-trusted senders, a denied sender could still trigger host-media attachment loading and cause readable local files to be returned through the outbound media path. ## Affected Versions This issue is known to affect OpenClaw 2026.4.9. Earlier versions were not confirmed during triage, so the advisory range is intentionally scoped to `>= 2026.4.9 < 2026.4.10`. ## Impact Affected deployments are those that both allow host read or filesystem root expansion at the global/agent level and rely on sender or group-scoped policy to deny `read` for some channel participants. In that configuration, the intended sender/group authorization boundary could be bypassed for outbound media reads, potentially disclosing host-local files readable by the OpenClaw process. The issue does not require treating the model prompt as the security boundary. The vulnerable behavior was a concrete policy enforcement mismatch: sender/group policy denied `read`, while the host-media read helper could still be installed without that sender context. ## Resolution Fixed in OpenClaw 2026.4.10 by PR #64459, commit `c949af9fabf3873b5b7c484090cb5f5ab6049a98`. The fix threads sender, session, channel, and account context into outbound media access resolution and intersects host-media read capability creation with the existing group tool policy for `read`. When a concrete sender/group override denies `read`, OpenClaw no longer creates the host `readFile` media capability. Additional attachment canonicalization hardening shipped in 2026.4.14, but the authorization bypass described here was fixed in 2026.4.10. ## Credit Thanks to @Telecaster2147 for reporting this issue.

Exploitation Scenario

An attacker with a less-trusted participant account in an OpenClaw channel — such as an external contractor or third-party integration explicitly denied `read` via `toolsBySender` — sends a crafted message that triggers the host-media attachment loading path. Because the host-media read helper is installed without evaluating the sender's policy context, it loads and returns files readable by the OpenClaw process through the normal outbound media response. The attacker receives file contents (e.g., `.env` files, SSH private keys, application configuration) in the channel response with no error indication. No specialized AI/ML knowledge is required beyond familiarity with OpenClaw's media attachment invocation syntax.

Weaknesses (CWE)

CWE-863 Incorrect Authorization Primary

CWE-863 — Incorrect Authorization: The product performs an authorization check when an actor attempts to access a resource or perform an action, but it does not correctly perform the check.

[Architecture and Design] Divide the product into anonymous, normal, privileged, and administrative areas. Reduce the attack surface by carefully mapping roles with data and functionality. Use role-based access control (RBAC) [REF-229] to enforce the roles at the appropriate boundaries. Note that this approach may not protect against horizontal authorization, i.e., it will not protect a user from attacking others with the same role.
[Architecture and Design] Ensure that access control checks are performed related to the business logic. These checks may be different than the access control checks that are applied to more generic resources such as files, connections, processes, memory, and database records. For example, a database may restrict access for medical records to a specific database user, but each record might only be intended to be accessible to the patient and the patient's doctor [REF-7].

Source: MITRE CWE corpus.