Off the top of my head, I know the cost is substantial if you want legit security. You have to perform updates monthly and everyone needs to be trained in each company how to use it, implement it, update it, and so on. 

Outside of cost - You have to worry about attackers that can use adversarial techniques to trick AI systems into making incorrect decisions. They can learn of mistakes and weakness which can be historically damaging to larger companies or to government agencies making use of it. 

I think the biggest risk are the people using it. From what I understand, you can train AI to detect things, prevent things, etc. but people can also train different AI to manipulate your own AI and seek out anything that is vulnerable.