Running with the Red Queen: Using game theory to improve cybersecurity in the AI era

by Eva Michely, Helmholtz Association of German Research Centres

edited by Gaby Clark, reviewed by Robert Egan

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources Running with the Red Queen: Using Game Theory to Improve Cybersecurity in the AI Era. Credit: CISPA/Chiara Schwarz In Lewis Carroll's "Alice Through the Looking Glass," the eponymous heroine encounters the Red Queen, who involves her in a frantic race. In this curious country, Alice learns, you have to run as fast as you possibly can only to remain in the same place.

Christoph Landolt draws on the image of this Red Queen's Race to describe the situation of software vendors in the AI era. As the cyber capabilities of frontier models continue to grow, the potential for attacks is multiplying at speed too. Using an approach from game theory, Landolt and his co-authors have calculated how the release of new AI models could be sequenced to give software vendors a strategic head start over cybercriminals.

Their paper is being presented today (Oct. 26) at the Conference on Game Theory and AI for Security ( GameSec 2026 ) in Ann Arbor, Michigan.

The fact that the capabilities of AI models multiply with every new release is a phenomenon that concerns Landolt. Landolt is a doctoral researcher in the research group of Faculty Professor Dr. Mario Fritz and observes the release of new frontier models with particular regard to their potential to compromise cybersecurity. "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," he says.

Shortly thereafter, in April 2026, Anthropic decided against a public release of Claude Mythos Preview because the model proved dangerously efficient at discovering previously unknown security vulnerabilities. "The question we're asking in our research," says Landolt, "is how does this dynamic change the ways in which we need to think about cybersecurity in the AI era?"

Traditionally, the practice of so-called responsible disclosure gives defenders a head start over cybercriminals. If, for instance, researchers uncover a vulnerability, they report it to the software vendor, who then develops and releases a patch. In this scenario, knowledge of the vulnerability becomes public along with the update, giving users a fair chance to protect themselves in time. However, publicly available frontier models make their offensive and defensive capabilities available to both sides at the same time.

"It's hard to say whether traditional responsible disclosure has become obsolete in the AI era," says Landolt. "But we know that the time it takes to go from a disclosed vulnerability to an attack has decreased drastically. As the zerodayclock.com website showed at one point, in 2018 it still took about 2.3 years as opposed to only about 10 hours in 2026."

This dynamic affects whether software vendors even consider fixing security vulnerabilities—and users consider installing patches—to be worthwhile at all. "If attackers can be faster at discovering new vulnerabilities than users can be at updating their software, the system begins to oscillate," explains Landolt. In the proverbial Red Queen's Race, software vendors and users then fall behind, no matter how fast they run: "The benefit of releasing a patch becomes negative, and the benefit of not releasing a patch also becomes negative. When we operate in this space, we're in a pickle no matter what we do."

Against this backdrop, Landolt advocates for releasing new frontier models incrementally and granting software vendors early access to them: "Through this head start, we can create an incentive for all patches to be already available when a new AI model becomes publicly accessible. This would allow me to protect my software effectively again."

In the above-mentioned case of Claude Mythos Preview, Anthropic opted for a similar approach: As part of its "Project Glasswing," Anthropic made the model available to select U.S. tech companies only. What Landolt and his co-authors Tobias Lorenz, Marta Kwiatkowska and Fritz are now proposing in "The Oracle's Gambit" potentially concerns the entire AI industry: Using an approach from game theory, the Stackelberg model, they have calculated how access to a new frontier model could be optimally sequenced from a cybersecurity perspective.

"We couldn't find any scientific literature that analyzes the impact of newly released AI models on cybersecurity," explains Landolt. "That's why we've developed a model that we can solve with an exact value, so we can get validation from the community and receive feedback on whether there are dynamics in practice that we haven't taken into account."

With their adaptation of the Stackelberg model, Landolt and his colleagues present a strategic two-level game involving three players. At the first level, the frontier lab itself acts as a player. It proposes a release policy for a new frontier model, determining the ways in which the AI will be made available. In doing so, it sets the rules according to which software vendors and cybercriminals compete against each other. At this second level, software vendors use the frontier model to develop patches for a set number of vulnerabilities, while cybercriminals use it to exploit these same vulnerabilities.

In this second game, software vendors are at a structural disadvantage. "The defender must not only ensure security but also availability: A vulnerability has to be detected, and a patch has to be built, tested and released," Landolt explains. "The attacker has it easier because they don't care if they cause a system crash." In addition, software vendors have to decide whether they will release a patch or withhold it. This is a strategic decision: Once a patch is released, it can be reverse-engineered and used for exploits by the opposing side.

As soon as this game ends, yielding an exact value, it is the frontier lab's turn again. It must now devise a release strategy that promises the greatest possible social welfare. In the model developed by Landolt and his colleagues, the frontier lab is motivated not only by altruism but also by economic incentives.

"Software vendors' disclosure policies generally assume that the developers themselves have a strong interest in protecting their users. With a frontier lab, however, the customer relationship is different, since the users of a software product are not necessarily customers of the frontier lab," Landolt says.

"Our model posits that the frontier lab will balance the opportunity costs of its lost market share against the welfare of the entire system. The idea behind this is that the frontier lab could suffer significant reputational damage if its models were used for malicious purposes." In "The Oracle's Gambit," release strategies geared toward the common good thus also counteract economic losses.

It is no coincidence that the researchers have chosen a game-theoretic approach to calculate a strategically sequenced release schedule for frontier models. "Cybersecurity and game theory are a common combination, based on theoretical considerations. 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," Landolt says.

However, this inclusion of a third party—and the dynamics they introduce into the interaction—posed logistical challenges for the researchers as they modeled the game's progression. "If we take into account all possible interactions between the attacker and the defender, our game becomes too complex, so we had to resort to simplified solution algorithms and assumptions. And even with this simplified model, we still needed two terabytes of RAM just to be able to solve the game once."

Christoph Landolt et al., The Oracle's Gambit: A Game-Theoretic Framework for Responsible AI Release (2026). DOI: 10.60882/cispa.33086237 Provided by Helmholtz Association of German Research Centres Who's behind this story? Gaby Clark MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile →

Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →