In this paper, we demonstrate the feasibility of a
competitive player using statistical learning methods to gain
an edge while playing a collectible card game (CCG) online.
We showcase how our attacks work in practice against the
most popular online CCG, Hearthstone: Heroes of World
of Warcraft, which had over 50 million players as of April 2016.
Like online poker, the large and regular cash prizes of
Hearthstone’s online tournaments make it a prime target for
cheaters in search of a quick score. As of 2016, over $3,000,000
in prize money has been distributed in tournaments, and
the best players earned over $10,000 from purely online
In this paper, we present the first algorithm that is able
to learn and exploit the structure of card decks to predict
with very high accuracy which cards an opponent will play in
future turns. We evaluate it on real Hearthstone games and
show that at its peak, between turns three and five of a game,
this algorithm is able to predict the most probable future card
with an accuracy above 95%. This attack was called “game
breaking” by Blizzard, the creator of Hearthstone.