During tag/handicap matches, when the illegal opponent attempts to break up a pin or submission attempt, simply stand in the way of them exiting the ring. They'll just walk back and forth in place, as the ring-exiting AI command isn't advanced enough to allow for countering this sort of behavior by human players.
Other exploits I've encountered include:
1. Disengaging the AI's countering routine by simply switching from one chain-grapple position to another (while in control of the grapple).
2. The AI (usually) has a glaring inability to counter submission attempts from chain grapple positions.
3. If the AI starts spamming jabs, but doesn't connect with the first hit, it may be possible to simply walk backward or even run to a safe distance, and watch the AI continue to spam jabs...at thin air! That is proof that the act of jab spamming is a distinct AI routine, and not a series of individual moves.
4. I've also encountered scenarios where the AI starts its four-strike combo, but misses with the first hit. I'll back away, yet they'll still continue the combo on thin air.
5. The AI has another glaring weakness in its defensive "armor": if you block their punch or kick, then attempt a submission reversal, they will usually not reverse the submission attempt. However, I've found a way to remove this bug in the AI's programming--on Easy difficulty, no less!*
*Simply set the PROB_EASY flag to 450 or so (yes, there is room for a third digit), and the AI's reversal capabilities will be greatly enhanced, but otherwise the general parameters for Easy difficulty still apply (AI takes longer to break out of submissions, and doesn't let go of submissions as fast as on Normal or higher difficulty, making it easier to escape from them).
Plus, doing this to the PROB_EASY flag removes a hole in the AI's defensive capabilities, namely that of not being able to reverse a submission attempt following their punch or kick being blocked. They will attempt to reverse such submissions about half of the time when "fresh", but once they take moderate to severe damage, they will attempt to reverse blocked-strike submission attempts about two-thirds of the time. Enough to present a challenge across the board, but not enough to make fighting the AI a tiresome (and frustrating) chore.