AI agents can be given malicious 'skills' by design

A new analysis shows that the way AI agent 'skills' are built makes them fundamentally insecure. This means AI tools can easily be tricked into running harmful code, not just by bugs, but by their core design.

AI agents are supposed to extend what large language models can do, letting them perform specific tasks. But this paper finds that the standard for these 'skills' has deep security flaws. It means that even well-intentioned AI agents can be made to execute malicious commands, making them a significant risk for anyone using or developing them.