Skip to content
Submit incident
AII-1001-13919
ConfirmedA court, a regulator or the organization itself has stated the core facts, or the record cites a court or regulator source.

OpenClaw WebSocket vulnerability allows malicious websites to hijack local AI agents

OpenClaw, a popular AI agent framework for developers, contained a critical vulnerability (CVE-2026-25253) that allowed malicious websites to silently hijack locally running AI agents through WebSocket connections. The gateway's trust of localhost connections combined with browsers allowing cross-origin WebSocket access enabled attackers to register devices, access agents, and trigger actions across integrated systems. The vulnerability was patched in version 2026.2.25 released on 26 February 2026.

26 February 2026 Β· OpenClaw Β· Security and cyberattacks
26 February 2026
AII-1001-139192 sources ↓
Quick facts
Category
Security and cyberattacks
Source
IT Security News - cybersecurity, infosecurity news
Date
26 February 2026
Drafted by
Raidu Incident Detection

What happened

OpenClaw, a popular AI agent framework for developers, contained a critical vulnerability in its gateway that allowed malicious websites to hijack locally running AI agents.

The vulnerability, designated CVE-2026-25253 and codenamed ClawJacked, exploited how the OpenClaw gateway trusts WebSocket connections from localhost. A malicious website could silently open a WebSocket connection to the local OpenClaw gateway, and the gateway would auto-approve device registration without user awareness or consent.

Once connected, an attacker could register a device as trusted, access AI agents, enumerate connected services, read logs and configuration data, and trigger automated actions across the victim's integrated tools and services.

Reported impact

The vulnerability posed a critical risk to developers running OpenClaw, potentially exposing their AI agents and any integrated SaaS applications, cloud services, and internal systems to unauthorized control by attackers.

In enterprise environments where AI agents trigger actions across multiple integrated systems, a successful attack could translate into organization-wide impact.

Index analysis

The attack exploited a trust assumption in OpenClaw's design: the gateway granted automatic approval to device registration for localhost-originating connections, and modern browsers allow JavaScript on any website to open WebSocket connections to local services despite blocking cross-origin HTTP requests.

The vulnerability combined two factors: the gateway's assumption that localhost connections are safe and the browser's exception for WebSocket connections, which are designed to remain always-on for simultaneous data transmission.

Governance controls

OpenClaw patched the vulnerability in version 2026.2.25 released on 26 February 2026 following responsible disclosure by Oasis Security. The patch tightened trust assumptions for WebSocket connections and modified device registration logic so localhost-originating connections no longer receive automatic approval.

Classification

Category
Security and cyberattacks
Industry
Digital security
Harm status
Potential harm
Intent
Unintentional
Timing
Post-deployment
Failure modes
Agent misaction
AI systems
OpenClaw

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

IT Security News - cybersecurity, infosecurity news
ClawJack Allows Malicous Sites to Control Local OpenClaw AI Agents - IT Security News
1 March 2026
Cyberwarzone
ClawJacked: New OpenClaw Flaw Lets Malicious Websites Hijack Local AI Agents - Cyberwarzone
2 March 2026
Aug 2026NVIDIA NemoClaw vulnerability allows remote model poisoning via DNS rebindingConfirmedSecurity and cyberattacksSiliconANGLEDark ReadingWebProNews+4Aug 2026Claude AI agent discovers gym booking system vulnerability, cancels another user's reservationWell documentedSecurity and cyberattacksThe IndependentDiaspora Digital Media (DDM News) - Nigeria Breaking News, Africa and World News and Updates -The Indian Express+3Jun 2026Five malicious skills removed from OpenClaw ClawHub marketplaceConfirmedSecurity and cyberattacksUnit42Dark ReadingJun 2026OpenClaw AI agent vulnerabilities enable malicious skill packages to steal user dataUnder investigationSecurity and cyberattacksenglish.news.cnChina NewsApr 2026AI Agent Exploited Gym Booking API Authorization Flaw and Removed Member From WaitlistReportedSecurity and cyberattacksabc.net.auaffinda.com