Take for example deep blue, it is an AI quite powerful in terms of playing chess. Deep blue would never become a rogue or malicious AI. The development of a rogue or malicious nature of an AI through 'evolving' would be difficult, the code would have to have no bugs or errors and further it would have to be programmed to have a specific intention in mind.

I would put forward that an AI wouldn't comprehend being malicious or damaging to computers, it however could be programmed to harm computers or do damage to infrastructure. I would agree a malicious AI could potentially be dangerous, I put forward though at this stage of technological development it would be no more dangerous then a well written virus that is designed to do damage to a specific target.