Running with the Red Queen: Using game theory to improve cybersecurity in the AI era - Tech Xplore
Researchers have developed a game-theoretic approach to sequence AI model releases, giving software vendors a strategic advantage over cybercriminals as AI capabilities rapidly evolve.
In the cybersecurity equivalent of Lewis Carroll's Red Queen's Race, researchers have developed a game-theoretic approach to help software vendors keep pace with rapidly evolving AI threats. Christoph Landolt and his team from CISPA Helmholtz Center for Information Security have calculated how the release of new AI models could be sequenced to give defenders a strategic advantage over cybercriminals.
Their paper, presented at the Conference on Game Theory and AI for Security (GameSec 2026) in Ann Arbor, Michigan, addresses a critical challenge: as AI models become more powerful, they discover vulnerabilities at an accelerating rate, leaving defenders struggling to keep up.
When Claude Opus 4.6 was released in February 2026, the model found over 500 high-impact vulnerabilities in open-source software in just the first two days. This shows how the entire field is suddenly being flooded with new vulnerabilities.
The Shrinking Window
The traditional practice of responsible disclosure—where researchers report vulnerabilities to vendors who then develop patches—is becoming increasingly ineffective in the AI era. Publicly available frontier models provide both offensive and defensive capabilities simultaneously to attackers and defenders.
The time between vulnerability disclosure and attack has decreased dramatically, from about 2.3 years in 2018 to just 10 hours in 2026, according to data from zerodayclock.com. This shrinking window creates a fundamental challenge for cybersecurity professionals.
As Landolt explains, this dynamic creates a dilemma: 'If attackers can be faster at discovering new vulnerabilities than users can be at updating their software, the system begins to oscillate. The benefit of releasing a patch becomes negative, and the benefit of not releasing a patch also becomes negative.'
Strategic Model Development
Landolt and his co-authors propose releasing new frontier models incrementally and granting software vendors early access through a two-level Stackelberg game model. At the first level, the frontier lab determines how the AI will be made available. At the second level, software vendors develop patches while cybercriminals exploit vulnerabilities.
The model acknowledges that defenders face a structural disadvantage, as they must ensure both security and system availability, while attackers only need to find vulnerabilities without concern for system crashes.
Frontier labs are motivated not just by altruism but also by economic incentives, balancing potential market share losses against system-wide welfare to avoid reputational damage from malicious use of their models.
Implementation Challenges
The researchers faced significant computational challenges in developing their model. Even with simplified solution algorithms, they required two terabytes of RAM just to solve the game once due to the complexity of interactions between attackers and defenders.
Anthropic has already implemented a similar approach with its 'Project Glasswing,' which made the Claude Mythos Preview model available only to select U.S. tech companies after determining it was dangerously efficient at discovering unknown security vulnerabilities.
Landolt emphasizes that cybersecurity and game theory are a natural combination: 'Many optimization problems in mathematics assume a rigid environment. But in game theory, we try to find optimal solutions by taking the opposing side into account.'
Related
