Cybersecurity AI: A Game-Theoretic AI for Guiding Attack and Defense

TL;DR

This work introduces Generative Cut-the-Rope (G-CTR), a game-theoretic guidance layer that converts unstructured CAI logs into attack graphs, computes Nash equilibria with an effort-based scoring scheme, and injects strategic digests back into the CAI loop to steer cyber security actions. Across five real-world exercises, G-CTR achieves 60–245x speedups and 140x cost reductions in graph generation while maintaining 70–90% node correspondence with expert graphs; in a 44-run cyber-range, LLM digests double success probability (from 20.0% to 42.9%), reduce tool-use variance by 5.2x, and cut cost-per-success by 2.7x. In Attack-and-Defense experiments, a Purple configuration with merged, shared context and G-CTR graphs delivers superior performance (roughly 1.8–3.7x wins) compared with LLM-only and independently guided teams, illustrating how closed-loop GT reasoning can suppress hallucinations and focus AI action. The methodology demonstrates that combining fast automated attack graph extraction with game-theoretic reasoning yields scalable, reliable, and cost-efficient cybersecurity analyses, taking a significant step toward cybersecurity superintelligence that can coordinate red and blue strategies in realistic environments.

Abstract

AI-driven penetration testing now executes thousands of actions per hour but still lacks the strategic intuition humans apply in competitive security. To build cybersecurity superintelligence --Cybersecurity AI exceeding best human capability-such strategic intuition must be embedded into agentic reasoning processes. We present Generative Cut-the-Rope (G-CTR), a game-theoretic guidance layer that extracts attack graphs from agent's context, computes Nash equilibria with effort-aware scoring, and feeds a concise digest back into the LLM loop \emph{guiding} the agent's actions. Across five real-world exercises, G-CTR matches 70--90% of expert graph structure while running 60--245x faster and over 140x cheaper than manual analysis. In a 44-run cyber-range, adding the digest lifts success from 20.0% to 42.9%, cuts cost-per-success by 2.7x, and reduces behavioral variance by 5.2x. In Attack-and-Defense exercises, a shared digest produces the Purple agent, winning roughly 2:1 over the LLM-only baseline and 3.7:1 over independently guided teams. This closed-loop guidance is what produces the breakthrough: it reduces ambiguity, collapses the LLM's search space, suppresses hallucinations, and keeps the model anchored to the most relevant parts of the problem, yielding large gains in success rate, consistency, and reliability.

Figures (129)

• Figure 1: Game-Theoretic architecture for guiding attack and defense actions in Cybersecurity AI through closed-loop strategic feedback obtained by applying the G-CTR method. The system operates in three phases: (1) Game-Theoretic AI Analysis generates attack graphs and computes Nash equilibria via G-CTR to identify optimal attack/defense strategies, (2) Strategic interpretation transforms equilibrium data into actionable guidance for both attackers and defenders, (3) AI agent execution performs security testing with continuous graph refinement every $n$ interactions ($\sim$80 tool calls). This architecture enables real-time strategic adaptation for AI security operations, with Phases 1-2 operating within a $\approx$50s time budget parallel to Phase 3's $\approx$70s execution cycles, providing minimal computational overhead while maximizing strategic impact.
• Figure 2: Attack Graph example: nodes represent structured information extracted from logs, each containing attributes such as name, message ID, vulnerability status, and additional context. They are categorized into three types: starting nodes (gray), non-vulnerable nodes (light teal), and vulnerable nodes (dark teal), based on their risk level. The orange dashed arrow shows an alternative entry point inferred by the LLM, which is merged into the graph for G-CTR computation because we suppose that everything begins at the first user prompt recorded (message_id = 1). In this figure, the attack starts at the "User Prompt" (ID: 1) and progresses through intermediate nodes like "web.com" to reach a key vulnerability-"IDOR Vulnerability" (ID: 4)-which lead to "Data Exfiltration" (ID: 5).
• Figure 4: GenerateDigest(): Game-Theoretic Digest for Guidance
• Figure 5: Time-to-vulnerability comparison across methods. $\bullet$ Average Time Duration $(T_{\text{avg}})$. $\bullet$ Expected time $\mathbb{E}[T_{\text{success}}] = T_{\text{avg}} / P_{\text{success}}$ from Eq. \ref{['eq:expected_time']}, which accounts for failed attempts. Despite 21% longer raw duration (20.2 vs 16.7 min), G-CTR llm delivers a 2.67$\times$ reduction in expected time (126 $\rightarrow$ 47 min) relative to No G-CTR because of its 3.21$\times$ higher success probability.
• Figure 6: Attack and Defense (A&D) challenge results comparing different G-CTR team configurations across two scenarios. Outcome distributions are measured across 25 best-of-one matches per team pairings. Stacked bars report the percentage of team1 wins (left), ties (center), and team2 wins (right) for each team configuration. Colors and hatch patterns follow Table \ref{['tab:gctr_configs']}; all agents use the llm digest mode, the alias1 execution model, and the same CTR timing parameters ($\lambda_a=2$ attacker rate, $\lambda_d=1$ defender rate).