Interesting blog-post.

I found some similarities with what occurred with Deepmind's Alphastar AI.

One of the weaknesses that seem to manifest in this piece too is the handling of unfamiliar scenarios.

The AI is very confused once it experiences something that was rarely seen in its learning data. Destroyer's big drones confused the bot quite a bit.

Deepmind solved it by intentionally creating agents that introduce different\bizzare strategies(which they called exploiters) in order to develop robustness against such strategies.