Agentic AI and the Cyber Arms Race

TL;DR

Agentic AI is poised to redefine cyber warfare and global politics by enabling autonomous, multi-agent cyber operations that automate offense and defense. The paper outlines a Centralized Reinforcement Learning Agent (CARL) architecture with specialized subagents (LREM, log agent, networking agent, vulnerability finder) and references real-world multi-agent platforms to illustrate practical trajectories. It analyzes how these capabilities could both erode and elevate existing power asymmetries, emphasizing co-evolutionary dynamics and the threat of adversarial AI, along with the potential diffusion of capabilities to mid- and small-state actors. The findings highlight significant strategic implications, including rapid capability proliferation, attribution challenges, and new forms of deterrence and instability that policymakers and defenders must address.

Abstract

Agentic AI is shifting the cybersecurity landscape as attackers and defenders leverage AI agents to augment humans and automate common tasks. In this article, we examine the implications for cyber warfare and global politics as Agentic AI becomes more powerful and enables the broad proliferation of capabilities only available to the most well resourced actors today.

Figures (1)

• Figure 1: Graph showing AI Red/Blue Agent coevolution. A low episodic return indicates the blue agent is performing well, while a high episodic return indicates the red agent is performing well. The vertical lines represent different training runs. In the first run, the red agent is trained without a blue agent. Next, the blue agent is trained against this version of the red agent, and so on. As can be seen, over multiple runs the agents can learn to adapt to changes in one another's abilities, effectively coevolving. Cyberwheel, the environment that generated this graph, is on https://github.com/ORNL/cyberwheel.