It's well known that it's easier to attack than defend, the reasons behind this were analytically deconstructed by Vukovic in his excellent book Art of Attack in Chess. Even Alpha-Zero is not immune. When it played millions of chess games against itself, it realized the importance of the initiative and prioritized the initiative so much that it rarely had to defend. When it had to defend tough positions where there was no chance of a counter-attack or active defense, it could collapse like anyone else. Here is an excellent example of this.
Like in the first game, Alpha-Zero can lose when forced to play a dubious opening.
In the following game, it was forced to play a Sicilian where White had the initiative. AlphaZero is stronger when it has the initiative, so it was not able to play into its strengths.
Since AlphaZero largely evaluates based on statistics (the computer equivalent of principles and patterns) and is weaker at calculation than Stockfish, it can sometimes fall into an opening trap and miss concrete details.
Another way that AlphaZero can lose is by overpressing in drawn endgames, a human-like mistake. Technique is not AlphaZero's strong suit, as in playing millions of games against itself, it quickly got more experience playing for mate in the middlegame, and due to that being its early strength, AlphaZero aimed for those types of positions more often. GM Mathew Saddler gave at least one example of AlphaZero blundering a drawn technical endgame because it had the worse chances in his book Game Changer: AlphaZero's Groundbreaking Chess Strategies and the Promise of AI, but the following 2 examples aren't like that. These examples show AlphaZero taking excessive risks in equal endgames trying to win. AlphaZero had less experience playing for a win in dry and equal endings compared to dynamic middlegames, so it didn't always know what types of risks would be objectively appropriate.
