But training on that particular sort of adversarial states should help against the human player which has learned the strategy, just like training on patch adversarial examples in vision helps against the same type of patches.

Of course if the adversarial policy is again allowed to find off-distribution states (by playing against the victim), it will certainly find ways to beat it, until the model is playing perfectly. (Emergent gradient obfuscation could also theoretically happen, but I don't know if it has been demonstrated to actually happen.)

We've apparently entered the stage where the deciding factor between who wins, man or machine, is just an arms race.

My understanding is that gwern above linked solid evidence in the paper for more search not being enough, as in, the model's evaluation NN is so way off target when searching, that realistic amounts of search don't help. Go seems to have many possible moves per position, so the search doesn't go very deep anyway.

Feel free to correct me if I'm wrong, it might be that I misremembered how AlphaGo-style systems work.

We're looking into doing our own adversarial training run to see if we can get to a point where the KataGo agent is robust. My personal suspicion given how difficult adversarial examples have been to eliminate in image classifiers and other ML systems is that although we'll be able to train particular vulnerabilities out of the system, there's still going to be a long tail of issues that can be automatically discovered and exploited.