Core War is a game where assembly language programs compete for dominance by attempting to crash one another within a virtual computer. In the newly announced method, LLMs drive the adversarial evolution within this environment.
LLMs Drive Adversarial Evolution in Core War, Observing Strategy Convergence and Self-Modifying Code
This article is a translation. Read the Japanese original
Rather than relying on traditional static benchmarks, the researchers observed a process of repeated adaptation to a growing pool of past opponents. This resulted in the emergence of more generalized strategies.
It was also confirmed that different code implementations converged toward similar high-performance behaviors. Furthermore, because code and data share the same address space, chaotic dynamics of self-modifying code occurred.
This environment is positioned as a secure sandbox for analyzing the evolution of AI agents in real-world adversarial settings, such as cybersecurity.
Sources: Digital Red Queen: Adversarial Program Evolution in Core War with LLMs (HN 126pt, 18 comments) (HN Search (backfill), 2026-01-09)