The Oracle's Gambit: A game-theoretic framework for responsible AI release

In an arXiv paper, the researchers propose using a two-level Stackelberg game to characterize AI model release decisions, analyzing how a lab should choose release timing when the capabilities of both attacker and defender derive from the same model. The paper points out that existing safety frameworks only regulate deployment thresholds, without modeling release timing; in this case the defender's first-mover advantage depends on the lab's release decision, and the question shifts from whether to release to how to release. In this framework, the lab commits to a release window, determining how both sides' capabilities evolve over time in downstream attack-defense competition. Model analysis shows that defender welfare depends on the capability gap rather than the common level; giving the same model to both sides may trap the defender in a Red Queen-style race, whereas releasing only to the defender in advance can create a protective gap, and the lab's optimal window must trade off welfare gains against the opportunity cost of delayed release. The paper accordingly argues that for dual-use models, the key lever is the order of access rather than deployment thresholds.