Critical OpenClaw Vulnerability Exposes AI Agent Risks

Dark Reading

by Jai Vijayan

March 2, 2026

AI-Generated Deep Dive Summary

A newly disclosed high-severity vulnerability in OpenClaw, a popular open-source AI agent tool, has been patched after researchers identified that malicious websites could hijack developers' AI agents without requiring plugins or user interaction. This critical flaw highlights the growing risks of integrating AI tools into organizational environments without adequate security measures. The vulnerability stemmed from OpenClaw's failure to differentiate between trusted internal connections and those originating from malicious websites running in a developer's browser, allowing attackers to gain full control over devices. The rapid adoption of OpenClaw since its November 2025 launch has made it the most starred project on GitHub, surpassing even React. The tool's popularity stems from its flexibility, local control, and ability to integrate with messaging apps, calendars, and developer tools for automation. However, this speed of adoption has also exposed organizations to new security risks, including vulnerabilities like CVE-2026-25253, which allowed token theft, as well as command injection and prompt injection attacks. Researchers found over 820 malicious "skills" on its marketplace, ClawHub, up from just 324 in early February. The latest vulnerability was particularly dangerous because it exploited a flaw in OpenClaw's assumption that all localhost connections could be trusted. Attackers could use JavaScript on compromised websites to open WebSocket connections directly to the AI agent gateway, bypassing authentication checks and enabling full device control. This issue underscores the urgent need for developers to prioritize security updates and exercise caution when using third-party plugins. The growing concern among stakeholders is the lack of governance around agentic AI tools like OpenClaw. As these tools become more integrated into organizational workflows, the risks of exploitation by

Verticals

securitytech

Originally published on Dark Reading on 3/2/2026