Skip to main content
Vulnerability Database/CVE-2026-100535

CVE-2026-100535: OpenClaw Privilege Escalation Vulnerability

CVE-2026-100535 is a privilege escalation flaw in OpenClaw that allows restricted senders to persist instructions executed by agents with broader capabilities. This post explains its impact, affected versions, and mitigation steps.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-100535 Overview

CVE-2026-100535 is a privilege escalation vulnerability in the OpenClaw npm package affecting versions >= 2026.4.5 and < 2026.8.1. The flaw allows session-derived text persisted to session memory to lose its originating requester's restrictions and untrusted provenance. A restricted external sender can plant instructions that a later unattended background (dreaming) agent executes with broader file and command capabilities. The issue is categorized as [CWE-863] Incorrect Authorization and is fixed in version 2026.8.1.

Critical Impact

Attackers with limited tool access can persist instructions into session memory that an elevated background agent later follows, enabling actions beyond the original turn's authority against files, commands, and services.

Affected Products

  • OpenClaw npm package openclaw versions >= 2026.4.5
  • OpenClaw npm package openclaw versions < 2026.8.1
  • Deployments with session-memory capture and dreaming features enabled

Discovery Timeline

  • 2026-09-26 - CVE-2026-100535 published to NVD
  • 2026-09-28 - Last updated in NVD database

Technical Details for CVE-2026-100535

Vulnerability Analysis

The vulnerability arises in OpenClaw's session-memory subsystem. When session-derived text is persisted, the system drops the metadata that marked the content as originating from a restricted or untrusted requester. Later, when the background dreaming agent retrieves that stored text, it treats the instructions as trusted input and executes them with its own broader capability set.

This authorization gap crosses a trust boundary between an interactive request handled with limited tools and an unattended agent holding file and command privileges. The flaw is a classic authorization confusion issue tracked under [CWE-863], where the enforcement point does not consult the correct subject when deciding what actions are permitted.

Root Cause

Session memory persistence does not retain provenance labels or capability constraints tied to the original sender. The dreaming agent operates on stored instructions without re-evaluating the originator's authority, so restrictions applied at admission time are not re-applied at execution time.

Attack Vector

An external sender with restricted tool access submits content that is admitted to the session. The content is captured into session memory and later selected for background processing by the dreaming agent. If the model follows the embedded instructions, the agent executes file, command, or service actions that the original requester could not have performed directly. Exploitation requires that the attacker-controlled content be captured, selected for later processing, and followed by the model.

No verified public exploit code is available. See the GitHub Security Advisory GHSA-62qm-6fjj-6g23 and the VulnCheck Privilege Escalation Advisory for technical details.

Detection Methods for CVE-2026-100535

Indicators of Compromise

  • Session memory entries containing imperative instructions, tool invocations, or command strings that originated from restricted external senders.
  • Dreaming agent audit logs showing file writes, command execution, or service calls that have no corresponding authorized interactive turn.
  • Unexpected changes to files or configuration shortly after a background processing cycle completes.

Detection Strategies

  • Audit stored session-memory records for provenance metadata and flag entries missing originator or trust labels.
  • Correlate background agent actions with the originating session turn to detect capability escalation across contexts.
  • Review installed openclaw package versions in package-lock.json and node_modules for versions in the vulnerable range.

Monitoring Recommendations

  • Log every dreaming agent tool invocation with references to the source memory entry and the original sender identity.
  • Alert on background agent file or command operations that target paths or services outside expected automation scope.
  • Monitor npm dependency manifests in CI/CD pipelines for openclaw versions below 2026.8.1.

How to Mitigate CVE-2026-100535

Immediate Actions Required

  • Upgrade the openclaw npm package to version 2026.8.1 or later across all deployments.
  • Inventory environments where session-memory capture and dreaming are enabled and prioritize those for patching.
  • Review and purge existing session-memory entries that were persisted while running a vulnerable version.

Patch Information

The vulnerability is fixed in OpenClaw 2026.8.1. Upgrade using npm install openclaw@2026.8.1 or update the project's package.json constraint accordingly, then regenerate the lockfile. Refer to the GitHub Security Advisory GHSA-62qm-6fjj-6g23 for release notes.

Workarounds

  • Disable session-memory capture or the dreaming agent until the upgrade is deployed.
  • Restrict the dreaming agent's file and command capabilities to the minimum required for its workload.
  • Reject or sanitize session content from restricted external senders before it is written to persistent memory.
bash
# Upgrade OpenClaw to the fixed release
npm install openclaw@2026.8.1
npm audit

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.