Nash proved that every zero-sum game has a perfect strategy called a Nash equilibrium. This is a possibly randomized strategy for both players such that given that one player follows the strategy, it doesn't pay the other player to deviate from her strategy. As an example, consider the game rock, paper, scissors. One Nash equilibrium here has both players choosing their signal randomly. If one player does that, there is no point for the other player to follow any other strategy (though it doesn't hurt).
Since Poker is a symmetric zero-sum game (both players are treated the same by the game), the value of the game is zero, and any Nash equilibrium on average results in zero expected profit for both players. According to Wikipedia, Cepheus is an approximation to such a strategy. As such, Cepehus' strategy may be randomized.
What about bluffing? Let's consider a variant of rock, paper, scissors in which players can bluff. Say each player writes their choice on a piece of paper, and then there is some structured dialog in which players announce their choice and offer the other player to capitulate. The random strategy is still perfect, in that if one player follows it, it doesn't pay for the other player to deviate from it.
While Cepehus (if it were perfect) shouldn't lose on the long run, it is not guaranteed to be a good strategy in practice, since it only promises you a small negative expected profit. Perhaps when playing against good human players it will just break even on average. That doesn't mean that humans are any better than it, just that they're not any worse. Another possibility is that while in theory the only guarantee is that it breaks even, in practice it wins against sub-optimal humans. This is an empirical question that the developers are testing right now through their website.