{"id":611,"date":"2026-04-21T00:46:51","date_gmt":"2026-04-21T00:46:51","guid":{"rendered":"https:\/\/quantusintel.group\/osint\/blog\/2026\/04\/21\/why-game-theory-failed-to-predict-the-two-biggest-ai-events-of-2026-and-my-framework-didnt\/"},"modified":"2026-04-21T00:46:51","modified_gmt":"2026-04-21T00:46:51","slug":"why-game-theory-failed-to-predict-the-two-biggest-ai-events-of-2026-and-my-framework-didnt","status":"publish","type":"post","link":"https:\/\/quantusintel.group\/osint\/blog\/2026\/04\/21\/why-game-theory-failed-to-predict-the-two-biggest-ai-events-of-2026-and-my-framework-didnt\/","title":{"rendered":"Why game theory failed to predict the two biggest AI events of 2026 \u2014 and my framework didn\u2019t"},"content":{"rendered":"

Author: Berend Watchus. Independent AI & Cybersecurity Researcher. April 21, 2026 \u00b7 Publication: OSINT\u00a0Team<\/p>\n

\"\"<\/figure>\n

Hypergame theory \u00b7 AI strategy \u00b7 Case\u00a0study<\/p>\n

Why game theory failed to predict the two biggest AI events of 2026\u200a\u2014\u200aand my framework didn\u2019t<\/h3>\n
\"\"<\/figure>\n
\"\"<\/figure>\n
\"\"<\/figure>\n

The Claude Code leak and Project Glasswing were not surprises to anyone using the right model. Here is how the hypergame framework called both events, and why classical game theory was structurally unable\u00a0to.<\/p>\n

On March 31, 2026, Anthropic accidentally shipped 512,000 lines of source code to the world. On April 7, 2026, they announced that their most capable model\u200a\u2014\u200aone whose offensive cybersecurity capabilities surprised its own engineers\u200a\u2014\u200awould not be released publicly. A private coalition of twelve organizations, drawn from Anthropic\u2019s own investor and customer base, received exclusive access\u00a0instead.<\/p>\n

One week after these two events, a peer-reviewed arxiv paper arrived: \u201cWhy Open Source? A Game-Theoretic Analysis of the AI Race\u201d by Mladenovic, Courville, and Gidel. It is a rigorous piece of work. It proves that finding a Nash equilibrium in the AI open-source decision is NP-hard. It proposes tractable Mixed-Integer Programming formulations. It is mathematically sound.<\/p>\n

It cannot explain either event. Not because it is bad research\u200a\u2014\u200ait is good research\u200a\u2014\u200abut because it uses the wrong model for the real\u00a0world.<\/p>\n

I want to be precise about what I mean by that, and I want to show my work. Because the framework I have been building since November 2024\u200a\u2014\u200aacross a series of papers and articles\u200a\u2014\u200anot only accommodates both events. It predicted the structural conditions that made both events not just possible but\u00a0likely.<\/p>\n

The question is not whether classical game theory is wrong. It is whether it models the game that is actually being played. In both cases, it does\u00a0not.<\/p>\n

The arxiv paper and what it gets\u00a0right<\/h3>\n

The Mladenovic et al. paper builds a clean formal model of the AI race. There are n players. Each chooses to open-source or close-source their model. Each gains from community contributions when they open-source (parameter \u03b4\u1d62) and gains from competitors when competitors open-source (parameter \u0394\u1d62\u2c7c). The utility function captures relative scientific progress in a winner-takes-all race.<\/p>\n

Within those assumptions, the analysis is tight. The NP-hardness proof reduces the Pure Nash Equilibrium problem to 3-SAT. The MIP formulation is a genuine contribution for computational tractability. The corollary about weaker players having incentives to deviate toward open-sourcing\u200a\u2014\u200abecause their impact on leaders is small while their community benefit is large\u200a\u2014\u200amaps plausibly onto real dynamics like Meta\u2019s LLaMA releases.<\/p>\n

The model earns its place in the literature. But it carries three structural assumptions that the real world violated in the same week the paper was published.<\/p>\n

The three assumptions that\u00a0failed<\/p>\n

1. All players share the same model of the game they are in.
2. Player capabilities are fixed parameters, not emergent surprises.
3. The action space is complete\u200a\u2014\u200aevery possible move is enumerated at the\u00a0start.<\/p>\n

Assumption one: shared game\u00a0model<\/h3>\n

Classical game theory, in the tradition running from Nash through the arxiv paper, requires what philosophers call a \u201ccommon knowledge of the game.\u201d All players agree on who is playing, what actions are available, and how payoffs are calculated. This is not a minor technical assumption. It is load-bearing. Remove it and the Nash equilibrium concept becomes undefined.<\/p>\n

Hypergame theory, introduced by Bennett and Dando in their 1979 analysis of the Fall of France, starts from the opposite premise. Players are not in the same game. They are in subjective games\u200a\u2014\u200aeach player\u2019s internal model of the strategic situation, which may differ radically from every other player\u2019s\u00a0model.<\/p>\n

The French High Command in May 1940 was not incompetent. They were playing a game\u200a\u2014\u200aMaginot doctrine, northern reinforcement\u200a\u2014\u200athat was internally coherent and rational. The Wehrmacht was playing a different game: the Ardennes option, which the French had assessed as physically impossible. Two rational actors, different games. The outcome was not an equilibrium. It was a catastrophe for the side that didn\u2019t know what game they were\u00a0in.<\/p>\n

My NPC and hypergame paper, written in November 2024, applied this framework directly to AI systems and the strategic environments they operate in. The core argument: the most dangerous strategic situations are not those where one actor is smarter than another. They are those where one actor doesn\u2019t know they are playing a different game entirely.<\/p>\n

Now look at March 31,\u00a02026.<\/p>\n

Anthropic was playing the game of production deployment. Standard CI heuristics. A known bug in an acquired runtime\u200a\u2014\u200aBun issue #28001, filed twenty days earlier, unfixed\u200a\u2014\u200acaused source maps to be served in production despite being disabled in configuration. Someone forgot to add *.map to\u00a0.npmignore. That is the game they were in: ship version 2.1.88, check the boxes,\u00a0deploy.<\/p>\n

The security research community was playing a different game simultaneously: open-source intelligence. Find the artifact. Mirror it before it disappears. Fork it. Analyze it. The moment the source map hit npm, the OSINT game was already won. Anthropic didn\u2019t know that game was being played in parallel.<\/p>\n

A Korean developer named Sigrid Jin was playing a third game: clean-room replication. Using oh-my-codex\u200a\u2014\u200aa competing AI\u200a\u2014\u200ahe rebuilt the entire Claude Code agent architecture in Python before sunrise. The result, instructkr\/claw-code, became the fastest-growing GitHub repository in history. 50,000 stars in two hours. Anthropic\u2019s DMCA campaign swept 8,000 repositories and couldn\u2019t touch it\u200a\u2014\u200aclean-room reverse engineering is established legal doctrine.<\/p>\n

Malware actors were playing a fourth game: opportunistic distribution. Fake repositories dressed as leaked Claude Code delivered Vidar v18.7 and GhostSocks to anyone who downloaded them.<\/p>\n

Four simultaneous games. No shared ontology. The arxiv model has one game. The hypergame model was designed for\u00a0this.<\/p>\n

The arxiv paper models the open\/closed source decision as a binary action in a shared strategic space. There is no variable for \u201carchitecture permanently escapes into clean-room replication overnight via a missing\u00a0.npmignore.\u201d That event is not an equilibrium deviation. It is a move in a game the model doesn\u2019t represent.<\/p>\n

Assumption two: stable capabilities<\/h3>\n

In June 2025, I published a paper on universal compression, the P vs NP divide, and what I called the hidden hand of code. The core argument: the P vs NP divide is not just a mathematical boundary. It is a structural feature of how all complex systems\u200a\u2014\u200abiological, computational, social\u200a\u2014\u200anavigate complexity. NP-hard problems don\u2019t get solved by real actors. They get routed around through heuristics, compression, lazy evaluation, and modular\u00a0reuse.<\/p>\n

Applied to the AI race: finding the optimal containment strategy for a complex production codebase is itself NP-hard. Real organizations don\u2019t compute it. They run heuristics: checklists, CI pipelines, code review processes, standard tooling. When a heuristic fails\u200a\u2014\u200aa known runtime bug, a missing gitignore entry\u200a\u2014\u200athe result is not a deviation from equilibrium. It is the predicted output of a system navigating an NP-hard problem with bounded rationality.<\/p>\n

The arxiv paper\u2019s parameters \u03b4\u1d62 and \u0394\u1d62\u2c7c are stable. They are set at the beginning of the game and remain fixed. This is standard in classical game theory. The problem is that the most consequential thing that happened in April 2026 was a capability that was not in anyone\u2019s parameter set.<\/p>\n

Claude Mythos Preview\u200a\u2014\u200athe model at the center of Project Glasswing\u200a\u2014\u200awas not designed as a cyberweapon. It was a general-purpose reasoning and coding model pushed to extreme performance. In internal testing, it autonomously identified thousands of zero-day vulnerabilities across every major operating system and browser, some over two decades old, surviving millions of automated tests and sustained human review. It generated working exploits without specific direction. Anthropic\u2019s own researchers described it with words like \u201cspooky\u201d and\u00a0\u201cscary.\u201d<\/p>\n

Nobody designed this capability in. It emerged from extreme performance in adjacent domains. The \u0394\u1d62\u2c7c parameter\u200a\u2014\u200athe benefit competitor j gains from player i\u2019s open-sourcing\u200a\u2014\u200adoes not have a slot for \u201cthis model accidentally became the most capable offensive cyber tool ever publicly documented.\u201d The arxiv model\u2019s utility functions are stable by construction. The real world is\u00a0not.<\/p>\n

My framework, drawing on the compounding argument from the Iceberg Series and the hidden hand of code paper, explicitly accounts for emergent capability that exceeds design expectations. The universe runs lazy evaluation. It doesn\u2019t brute-force solutions. Neither do the actors in it. What compounds through iterative cycles of training and feedback produces outputs that weren\u2019t in the original specification\u200a\u2014\u200aand the strategic landscape has to be modeled accordingly.<\/p>\n

Assumption three: complete action\u00a0space<\/h3>\n

The arxiv model\u2019s action space is {0, 1}\u200a\u2014\u200aclosed source or open source. In the continuous extension, it becomes [0, 1]\u200a\u2014\u200apartial open-sourcing. This is a reasonable simplification for analyzing the open\/closed source decision. But it means the model cannot represent actions that weren\u2019t enumerated at the\u00a0start.<\/p>\n

\u201cPlayer consolidates access by forming a private coalition drawn from its own investor base, granting exclusive use of an emergent offensive capability to twelve organizations under a safety framework, thereby exiting the competitive race by becoming its gatekeeper\u201d is not in {0,\u00a01}.<\/p>\n

This is what Project Glasswing is. A private coalition including Amazon, Apple, Microsoft, Google, Nvidia, CrowdStrike, and JPMorgan Chase received exclusive access to Mythos under a cybersecurity initiative backed by $100 million in usage credits. There was no public process for that selection. No regulatory oversight. No democratic input. A private company made a unilateral decision about who controls the most capable offensive cyber tool ever publicly documented and named the decision a safety initiative.<\/p>\n

From the arxiv model\u2019s perspective, this is simply not representable. The game assumes players remain in the race, choosing open or closed source based on their position and community benefit. A player who achieves capability so decisive that they withdraw from the competitive field and restructure who can play\u200a\u2014\u200athis is not a Nash deviation. It is a move in a game the model doesn\u2019t\u00a0contain.<\/p>\n

My Iceberg Series Part 3, published on March 7, 2026\u200a\u2014\u200athirty-one days before Glasswing\u200a\u2014\u200aargued this structural outcome directly. The M2C (military-to-civilian) pipeline pattern, documented across GPS, ARPANET, the digital camera, and Silicon Valley\u2019s own origin story, predicts that when capability becomes strategically decisive, access is consolidated before it is democratized. GPS precision existed in operational military use from 1973. The public received a deliberately degraded version\u200a\u2014\u200aSelective Availability\u200a\u2014\u200auntil the year 2000. The paper stated: any person in a position of strategic responsibility who knowingly allowed a significant AI capability gap to develop while possessing the resources to prevent it would be committing dereliction of duty. Glasswing is Selective Availability for AI, implemented in weeks rather than decades because the capability emergence happened faster than policy could\u00a0respond.<\/p>\n

Glasswing does not prove the Iceberg argument. It is consistent with it. That is the appropriate epistemic claim\u200a\u2014\u200aand it is a strong\u00a0one.<\/p>\n

There is one more connection worth making explicit. My NPC paper, written in November 2024, proposed that high-risk AI applications should be deployed as exclusive in-house systems, controlled by vetted organizations, for reasons of security, accountability, and sovereignty. Section 7 of that paper outlines the governance architecture in detail: encrypted internal networks, custom security protocols, full control over model behavior, auditability, human oversight, reduced dependency on third-party providers.<\/p>\n

Glasswing is structurally identical to this proposal. The governance intent diverged\u200a\u2014\u200athe paper argued for it as responsible practice; Glasswing implements it as strategic consolidation\u200a\u2014\u200abut the structural outcome is the one I described sixteen months earlier. The in-house coalition model, the controlled access framework, the vetted partner list: I was arguing for this architecture in a different register. The same architecture arrived, for different reasons, on April 7,\u00a02026.<\/p>\n

Why the hypergame model is structurally superior<\/h3>\n

I want to be careful here. The claim is not that the arxiv paper is wrong. It is that it models a simplified version of the real game, and the simplifications are the ones that matter most for prediction.<\/p>\n

Three specific failure\u00a0modes:<\/p>\n

The shared ontology\u00a0failure<\/h3>\n

Classical GT requires a common knowledge of the game. The Claude leak involved four simultaneous subjective games with no shared ontology. The hypergame model\u200a\u2014\u200aintroduced by Bennett and Dando in 1979, applied to AI systems in my NPC paper in 2024, formalized computationally by Trencsenyi in December 2025\u200a\u2014\u200adoes not require shared ontology. It models each player\u2019s subjective game independently and asks: what happens when rational actors operating in different games interact? The answer is not equilibrium. It is structured surprise.<\/p>\n

The stable parameter failure<\/h3>\n

Classical GT\u2019s utility functions are fixed at design time. The P vs NP paper establishes that real systems route around NP-hard optimization rather than solving it\u200a\u2014\u200awhich means real capabilities emerge from heuristic processes that can produce outputs exceeding their own specifications. Mythos\u2019s offensive capability was not in anyone\u2019s \u0394\u1d62\u2c7c. The hypergame framework, combined with the compounding argument from the Iceberg Series, explicitly accounts for emergent capability that changes the strategic landscape mid-game.<\/p>\n

The closed action space\u00a0failure<\/h3>\n

Classical GT enumerates actions at the start. Glasswing\u2019s structural move\u200a\u2014\u200abecoming the gatekeeper of the game rather than a player within it\u200a\u2014\u200awas not in the action space. The hypergame model does not require a closed action space. Actors can discover moves that were not enumerated at design time. This is, in fact, the defining feature of hypergame advantage: the actor who finds a move outside the opponent\u2019s game model wins by a margin that classical GT cannot compute, because it is literally infinite\u200a\u2014\u200aa move the opponent cannot even perceive.<\/p>\n

The chronological record<\/h3>\n

This is not retrospective pattern-matching. The dates\u00a0matter.<\/p>\n

November 2024\u200a\u2014\u200aNPC \/ hypergame paper:<\/strong> Framework for overlapping subjective games, emergent strategic deception, in-house coalition governance model. Sixteen months before Glasswing.<\/p>\n

June 2025\u200a\u2014\u200aP vs NP \/ hidden hand paper:<\/strong> Real systems route around NP-hard containment via heuristics. Emergent capability exceeds design specifications. Nine months before the\u00a0leak.<\/p>\n

February\u2013March 2026\u200a\u2014\u200aIceberg Series 1\u20133:<\/strong> M2C pipeline predicts consolidation before democratization. Selective Availability argument. Gap they are required to keep. Thirty-one days before Glasswing.<\/p>\n

March 31, 2026\u200a\u2014\u200aClaude Code leak:<\/strong> Four simultaneous games. Overnight clean-room replication. Heuristic failure over NP-hard containment. No shared ontology.<\/p>\n

April 7, 2026\u200a\u2014\u200aProject Glasswing:<\/strong> Private coalition. Selective Availability. Player captures the game. M2C pipeline confirmed structurally.<\/p>\n

April 17, 2026\u200a\u2014\u200aarxiv game theory paper:<\/strong> Rigorous formal model of the open\/closed source decision. NP-hardness proof. Cannot represent either\u00a0event.<\/p>\n

What this means for how we model AI\u00a0strategy<\/h3>\n

The practical implication is not that classical game theory is useless. It is that it should be used for what it is: a model of stable, bounded, consensually-framed strategic interactions among actors who agree on the rules. It is excellent for those situations. Regulatory design, in particular, benefits from tractable equilibrium analysis of the kind the arxiv paper provides.<\/p>\n

But the AI race as it is actually unfolding is not that situation. The actors are not all in the same game. Capabilities are emerging that weren\u2019t in the original specifications. The action space is being rewritten in real time by actors who discover moves that weren\u2019t enumerated. The most consequential strategic outcomes in April 2026 happened in the space between the games the actors thought they were\u00a0playing.<\/p>\n

Hypergame theory was built for exactly this. Bennett and Dando built it to explain the Fall of France\u200a\u2014\u200aa catastrophic outcome that classical GT assessed as \u201cFrench incompetence\u201d and that hypergame theory correctly identified as rational actors in different subjective games. Vane deployed it at General Dynamics for over twenty years. Trencsenyi formalized its computational tractability in December 2025. My research program has been applying it to the AI race, to emergent AI capability, to the M2C pipeline, and to the governance question since November\u00a02024.<\/p>\n

The Claude Code leak and Project Glasswing are not edge cases in the AI race. They are the AI race, as it is actually unfolding. Any model that cannot represent them is a model of a different, simpler race\u200a\u2014\u200aone that is not happening.<\/p>\n

The French High Command was not incompetent. They were rational actors in a game that had already been superseded by a game they could not perceive. The question for AI researchers, policymakers, and security professionals is which game they are in right now\u200a\u2014\u200aand whether the model they are using can even show them the\u00a0answer.<\/p>\n

Conclusion<\/h3>\n

I do not write this to dismiss the arxiv paper. I write it because the field needs both kinds of models\u200a\u2014\u200athe tractable formal model for the game we can see, and the hypergame model for the game that is actually being played. The former tells you where the equilibria are if everyone agrees on the rules. The latter tells you what happens when they don\u2019t\u200a\u2014\u200awhich is, historically, where the decisive outcomes\u00a0occur.<\/p>\n

Two landmark events in April 2026 tested both models simultaneously. The formal game theory model was structurally unable to represent either. The hypergame framework, built across a series of papers between November 2024 and March 2026, accommodated both and had predicted the structural conditions for both before they occurred.<\/p>\n

That is not coincidence. It is what the right model\u00a0does.<\/p>\n

Related work by the author<\/strong> ChatGPT-Powered NPCs: AI-Enhanced Hypergame Strategies for Games and Industry Simulations (Zenodo, November 2024 \/ July 2025) \u00b7 From Chaos to Efficient Computing: Universal Compression, the P vs NP Divide, and the Hidden Hand of Code (Preprints.org, June 2025) \u00b7 The Chatbot LLM Asymmetry (OSINT Team, February 2026) \u00b7 The AlphaGo Moment for NPCs Happened in 2023 and Everyone Laughed (OSINT Team, March 2026) \u00b7 The Gap They Are Required to Keep\u200a\u2014\u200aIceberg Series Part 3 (OSINT Team, March 2026) \u00b7 The Claude Code Leak: What\u2019s Now Publicly Usable and Abusable (System Weakness, April 2026) \u00b7 Glasswing Is the Confirmation: The Manhattan Project for AI Arrived on April 7, 2026 (OSINT Team, April\u00a02026)<\/p>\n

From Chaos to Efficient Computing: Universal Compression, the P vs NP Divide, and the Hidden Hand of Code<\/a><\/p>\n

\"\"
https:\/\/www.preprints.org\/manuscript\/202506.2408<\/a><\/figcaption><\/figure>\n
\"\"<\/figure>\n