One of the most intractable debates in IR revolves around the severity and frequency of the security dilemma. Offensive realists argue that states are compelled to make worst-case assumptions about each other’s intentions, which yields inexorable competition and conflict even between mutually-benign actors. Yet others have argued that rational benign states should always be able to find cooperative signals that are costly enough to be credible, but not too costly to risk sending. If true, this should alleviate the security dilemma and facilitate cooperation, even under high initial distrust. However, there is little empirical work on interstate reassurance and the conditions under which mutually-benign actors can build trust. We address this gap using laboratory experiments to test Andrew Kydd’s canonical model of the security dilemma. We find strong support for the directional effects of the hypothesized signaling mechanisms. However, the frequency of cooperation is significantly lower than the model predicts, and the feasibility of reassurance is highly sensitive to the degree of prior trust. This implies that although reassurance can mitigate the security dilemma, offensive realism may still capture important psychological mechanisms that impede interstate cooperation.