OpenClaw Flaw Enables Agent Hijacking
HighExecutive Summary
A high-severity vulnerability, dubbed ClawJacked (CVE-2026-25253), was discovered in the OpenClaw AI agent framework. The flaw allowed a malicious website to silently brute-force access to a local instance, leading to token theft, full control, and potential data exfiltration; a patched version is now available.
Vulnerability Details
- Affected Product: OpenClaw (also known as Moltbot and Clawdbot) versions prior to 2026.2.26.
- Identifier: CVE-2026-25253
- CVSS Score: 8.8 (High)
- Exploitation Status: Proof-of-concept exists; no evidence of active exploitation in the wild.
Risk & Impact
- Triage: Urgent. The vulnerability allows for remote code execution and full system compromise with minimal user interaction.
- Attack Vector: A user running a vulnerable OpenClaw instance visits a specially crafted malicious website. The site uses a WebSocket connection to connect to the local OpenClaw gateway, which improperly trusts localhost traffic, allowing the attacker to brute-force the password and gain control.
- Ease of Exploit: Simple. The attack can be triggered by a single click on a malicious link, leading to compromise in milliseconds.
Action Plan
- Immediate Action: Upgrade to OpenClaw Version 2026.2.26 or later.
- Workaround: If immediate patching is not possible, avoid browsing untrusted websites while logged into the OpenClaw Control UI and consider using isolated browser profiles. It is also recommended to rotate any tokens or credentials used with a vulnerable version.
- Detection: Monitor logs for any unauthorized configuration changes or unexpected command executions.
Relevant professional terms
- AI Agent
- An autonomous program designed to perceive its environment, make decisions, and take actions to achieve specific goals without direct human intervention. In cybersecurity, they can act as automated defenders that monitor networks and respond to threats.
- Brute-force Attack
- A cyberattack method that uses trial-and-error to guess credentials like passwords or encryption keys. Attackers systematically try all possible combinations, often using automated tools, until the correct one is found.
Source: BleepingComputer
