AI is dangerous when it is developed into an autonomous weapon. But ultimately this is the intent of the builders of such systems. Really they are nothing more and nothing less than a type of organic weapon, like bioweapons. They can't kill everyone, and they will have weaknesses that especially human controlled smart robots will be able to find, exploit, and even, potentially, these could be purely viral responses that cause the AI weapons to turn on each other.

People think there is no way to deal with this threat, don't realise that malicious AI can be attacked, precisely as this story exemplifes, by other AI's programmed to target these other AI's. Even if they make them pass turing tests by restricting them to human languages, they still will have signatures that cannot be completely hidden in their way of communicating. Your personal way of writing, given sufficient quantity of text can identify you as surely as a fingerprint. The machines will be no different, and the more complex their activities become, the more distinctive they will be.