AI's pitting themselves against each other
Usually goes quite wrong...
It is possible to imagine scenarios where successful defence ramps up attack to epic proportions. Just like automated trading systems can cause insane and unbalanced share price changes.
Closer to the real world...
The challenge I fear is that finding these AI's is likely to require more cooperation between providers in sharing some level of traffic data between themselves to identify and thwart the miscreants. Trying to defend at your own endpoint will become more like throwing a rock into a pond.
AI for the low and slow probing is probably more effective at looking like "real" traffic too, also making it harder to detect and mitigate against. Security at application level potentially becomes a far more important level of defence (i.e. not assuming the "techies" can mitigate for me). I don't know many shops that look at application behaviour as an insight into trouble brewing...