Security analysis of OpenClaw: researchers test the AI coding agent against 47 attack scenarios. 🦞🛠️👀👨🏻‍💻📜

A group of security researchers from Shandong University (China) published their security analysis of the OpenClaw agent earlier this month. The authors tested 47 scenarios against the AI agent across six major attack categories (combining the MITRE ATLAS and ATT&CK frameworks).

According to the research, OpenClaw is quite vulnerable by default: it relies on the underlying language model for security controls and safety. Defense success rates vary depending on the model, ranging from about 17% to 83%.

The proposed solution is... humans! :) Human-in-the-Loop (HITL) is a defense mechanism the authors recommend implementing. What do you think?

This is one of the first public security assessments of OpenClaw, and there are many very interesting insights and ideas. Definitely a must-read if you’re into modern AI cybersecurity or want to get into it soon.

More details:

Don’t Let the Claw Grip Your Hand: A Security Analysis and Defense Framework for OpenClaw [PDF]: https://arxiv.org/abs/2603.10387

Code and samples [Github]: https://github.com/S2yyyy/OpenClaw-Analysis