The Oracle's Gambit: A Game-Theoretic Framework for Responsible AI Release

Timing of AI release matters more than the binary decision to release or withhold. The paper argues that existing safety frameworks focus only on whether to deploy an AI model, but the real strategic question is when to release it. The lab's decision about the release window—not just the yes/no choice—fundamentally shapes outcomes for both defenders and adversaries.

Giving the same AI model to both defenders and adversaries can backfire. When defenders and adversaries access identical models simultaneously, they enter a "Red Queen's race" where neither side gains a meaningful advantage, potentially leaving systems more vulnerable. The capability gap between defender and adversary matters more than the absolute capability level they both possess.

Sequencing access creates protective advantages for defenders. Pre-releasing an AI model to defenders before making it publicly available creates a "protective gap" that gives defenders time to prepare defenses. This sequencing strategy is more effective than simply withholding models, offering a practical middle ground between full transparency and complete secrecy.

Labs must balance security benefits against the costs of delayed release. The optimal release window trades off the security gains from giving defenders a head start against the opportunity costs of delaying public access to beneficial AI capabilities. This suggests responsible AI release requires deliberate strategic planning rather than immediate deployment.

For dual-use AI systems, access sequencing is the key lever, not deployment thresholds. The paper reframes how we should think about controlling dual-use AI technologies—the focus should shift from deciding whether to deploy to carefully managing who gets access when.