Game Theory Sequences AI Model Releases to Improve Cybersecurity

A new game-theory framework proposes sequencing the release of advanced AI models to give software vendors a strategic head start over cybercriminals. Christoph Landolt, a doctoral researcher in Mario Fritz’s research group, presented the study, The Oracle’s Gambit, with co-authors Tobias Lorenz, Marta Kwiatkowska and Fritz on Oct. 26 at the Conference on Game Theory and AI for Security in Ann Arbor, Michigan. The work argues that frontier models can accelerate both offense and defense, making release timing a core cybersecurity lever. The paper draws on the Red Queen’s Race from Alice Through the Looking Glass, where participants must keep running only to stay in place. In software security, that race has intensified as models improve their ability to discover vulnerabilities. Landolt cites Claude Opus 4.6, released in February 2026, which reportedly found more than 500 high-impact vulnerabilities in open-source software within two days. In April 2026, Anthropic chose not to publicly release Claude Mythos Preview after it proved unusually capable at uncovering unknown flaws, instead making it available to selected U.S. technology companies through Project Glasswing. These cases illustrate why traditional responsible disclosure is under pressure: public AI models can give attackers and defenders similar capabilities at the same time, while the window between vulnerability disclosure and exploitation has collapsed. The study references data from zerodayclock.com showing that the time from disclosure to attack fell from about 2.3 years in 2018 to about 10 hours in 2026. To address the dynamic, the researchers propose that AI labs release frontier models incrementally and grant software vendors early access. If patches are prepared before broad public release, defenders regain an incentive to maintain security updates. The Oracle’s Gambit models this process as a two-level Stackelberg game. At the first level, the frontier lab sets the release policy, including who may use the model and when. At the second level, software vendors and cybercriminals compete by using the model to fix or exploit a set of vulnerabilities. Defenders remain disadvantaged because they must preserve system availability and test patches, whereas attackers can succeed by causing damage alone. The framework also treats the frontier lab as a rational actor balancing economic incentives and systemic harm. Because software users are not necessarily customers of the AI lab, the model assumes the lab weighs lost market share from restricted access against reputational and societal damage if its model is misused. The authors said finding exact solutions required simplified assumptions and substantial computing resources, including two terabytes of RAM to solve one version of the game. Their central policy takeaway is that coordinated, staged model access could make patch release and user adoption meaningful again in an era where AI-assisted exploitation moves faster than traditional security response.