Evidence on the effect of extreme weather events or natural disasters on the risk of armed conflict is mixed. While it could be the case that one of the competing arguments is "true" and others are not, this paper examines another possibility: disasters can have heterogeneous effects on conflict risks and, consequently, systematically make empirical evidence mixed. Specifically, I develop a model of armed conflict that focuses on the role of political groups' asymmetric exposure to disasters. It presents two contrasting equilibria arising from a single theoretical framework: disaster-induced power shifts. In the first, a group attacks its rival after a disaster opportunistically if the former incurred disproportionately smaller damage and finds the rival temporarily vulnerable. In the second equilibrium, a player attacks the other preventively before a disaster occurs when the former is inherently vulnerable to climate anomalies and anticipates that it will be disadvantaged once a disaster erupts. These equilibria provide a novel interpretation of why the empirical evidence on climate conflict seems ostensibly mixed: (i) in the first case, the presence of a disaster event is positively correlated to conflict, whereas (ii) they are negatively associated in the second. This observation suggests that some empirical frameworks could underestimate the causal effect of disasters on conflict risks. I also connect the theory to the data of intrastate conflicts and droughts in African countries.