I don't have enough reputation to comment but accepted answer is incorrect. It cannot be AUC because AUC is non-negative.

Many things have changed in XGBoost since this question was posted but the values of the leaves are their weights and are actually part of the final prediction resulting from this tree. Final prediction is calculated via log-odds (which are additive) and are calculated as follows:

log_odds = leaf_value_tree_0 + leaf_value_tree_1 +...+ logit(base_score)

where leaf_value_tree_x is a leaf value in a tree x for given instance and base_score is initial probability of a target.

From this you can calculate probability (or score) as:

p = expit(log_odds)

Therefore log_odds and leaf_value_tree_x can be negative, while final probability p is between 0 and 1, as it should be.

More details on how leaf weights are calculated can be found in the documentation