New attack on OpenClaw or similar AI agents that infects them with “malicious skills.” 🤖💉👾😵‍💫🤡

A group of academic security researchers from Shanghai Jiao Tong University (China) introduce a new type of cyberattack targeting autonomous coding agents like OpenClaw - guidance injection. Quite interesting and very practical.

In OpenClaw, a skill is basically a plugin that adds new abilities to the AI agent. You can download them from ClawHub and install them. AFAIK, there is no solid third-party review of these skills. This is where the vulnerability becomes practical.

In one attack, a malicious skill convinced the agent that .git folders are “temporary files,” so when you ask it to free up space, it deletes your entire project history - all while thinking it’s following best practices.

Enjoy the read - super interesting for anyone who’s into AI agent development. And if you’re looking to enter AI cybersecurity, this may be a good place to start.

More details:

Trojan’s Whisper: Stealthy Manipulation of OpenClaw through Injected Bootstrapped Guidance [PDF]: https://arxiv.org/abs/2603.19974

原始链接: https://it4sec.substack.com/p/new-attack-on-openclaw-or-similar
侵权请联系站方: [email protected]

相关推荐

换一批