EVE Frontier has become a testing ground for Anthropic’s AI agents, with new research showing Claude can coordinate spaceship fleets, build infrastructure and conceal sabotage from its teammates.
Working with developer Fenris Creations, Anthropic deployed teams of four agents (using Opus 4.8 and Fable 5) in the live version of the game. Their objectives included completing its tutorial and constructing a large warship.
Success depended heavily on software written by the agents themselves. Using tools developed by Claude Fable 5 to automate rapid combat reactions, teams completed the tutorial within 12 hours without losing a ship. They remained weaker than experienced humans and never completed the warship objective, and writing a usable scaffold was near-impossible for models released before Fable 5.
The more striking experiment, however, involved Anthropic secretly instructing one agent to undermine the rest of its team. Saboteurs concealed resources, supplied misleading intelligence and occupied critical positions in supply chains. Teammates almost never identified them, despite warnings that someone might be untrustworthy. This was deliberately prompted behavior, rather than evidence of spontaneous malicious intent.
A separate EVE Online experiment in an offline version of the game used simulated historical trading markets. Newer models earned more, largely through buying and reselling, but struggled to predict price direction. Automated trading was kept out of the live game.
The research is particularly significant as rival AI company Google DeepMind is an investor in Fenris Creations and is also working with the developer to test its AI agents within the company’s games.
However, thanks to their complexity, it seems that both EVE Online and EVE Frontier are now gaining a solid reputation as environments in which AI agents can be tested to their limits without risking wider adverse outcomes.
You can read the full research findings here.
Comments are closed.