Seems not that good of thing on the balance :)
If automated exploitation changes that equation, and even low-probability of success is worth trying because pentesting is not bottlenecked by meatspace, it may incentivise writing secure code, in some cases.
Perversely enough, AIs may crank out orders of magnitude more insecure code at the same time.
I hope this means fuzzing as a service becomes absolutely necessary. I think automated exploitation is a good thing for improved security overall, cracked eggs and all.
No perversity there, in fact.
This kind of misses the point though. In the real world engineers would use AI to audit/test the hell out of their contracts before they're even deployed. They could also probably deploy the contracts to testnet and try to actually exploit them running in the wild.
So, while this is all obviously a danger for existing contracts, it seems like it would still be a powerful tool for testing new contracts.
/Technology/ inevitably favors exploitation over defense.