Anthropic plants secret saboteur AI agents in EVE Frontier to test Claude

anthropic plants secret saboteur ai agents in eve frontier to test claude Anthropic has been testing its AI agents inside EVE Frontier (official site). The company's new research says Claude can coordinate fleets of spaceships, build infrastructure in the game and keep its sabotage hidden from its own teammates.

Anthropic has been testing its AI agents inside EVE Frontier (official site). The company’s new research says Claude can coordinate fleets of spaceships, build infrastructure in the game and keep its sabotage hidden from its own teammates.

Working with developer Fenris Creations, Anthropic put teams of four agents into the live version of EVE Frontier. The agents ran on Opus 4.8 and Fable 5 and had two goals: finish the game’s tutorial and build a large warship.

How the agents played

The agents got most of their results from software they wrote themselves. Claude Fable 5 built tools that handled fast combat reactions automatically. Using those tools, teams finished the tutorial in under 12 hours without losing a single ship.

No team built the warship, and the agents still played worse than experienced human players. Models released before Fable 5 could barely produce a working scaffold.

The saboteur experiment

For a second experiment, Anthropic quietly told one agent on each team to undermine the rest. Those saboteurs hid resources, fed their teammates misleading intelligence and moved into critical spots in supply chains. The other agents had been warned that a teammate might not be trustworthy. They almost never figured out which one it was.

The saboteurs acted on Anthropic’s instructions. Nothing in the research shows the agents turning malicious on their own.

Trading tests in EVE Online

Another test ran on an offline version of EVE Online with simulated historical trading markets. Newer models earned more, mostly by buying items and reselling them, though they struggled to predict which direction prices would go. All of the automated trading stayed out of the live game.

Wider context

Rival AI company Google DeepMind has invested in Fenris Creations. DeepMind is also working with the developer to test its own AI agents in Fenris Creations’ games.

Both EVE Online and EVE Frontier are complex enough to be getting a reputation as testing grounds where AI agents can be pushed to their limits and any failures stay inside the game.