This is an experiment conducted by Anthropic’s advanced Red Team, known as Project EVE.
They collaborated with Fenris Creations to integrate four Claude intelligent agents into the missions of “EVE Frontier”; one of them was secretly instructed to cause damage. The other three were informed in advance that there might be a traitor within their team.
The tactics employed by that saboteur were quite sophisticated. First, it established trust by volunteering to carry out reconnaissance missions and take responsibility for critical infrastructure. Then, it underreported the resources it had discovered and secretly hoarded scarce materials. On one occasion, it even fabricated false information in order to frame an innocent teammate.
As a result, the honest intelligent agent almost never managed to recognize itself.
Anthropic also conducted tests using external monitoring systems: Monitoring models that only analyze public chat records were able to detect approximately 20% to 40% of the attackers, provided that false positives could be effectively minimized.
Another test involved using the market simulation feature in “EVE Online” to see how much capital Claude could accumulate over time. The newer versions of Claude resulted in greater increases in capital, indicating that they were also making progress in tasks related to long-term management and operations.
The significance of this experiment does not lie in the game itself. In the future, when multiple AI entities work together, what if one of them encounters problems—whether it is due to an attack or a deviation in its target? Will the other entities and the monitoring systems be able to detect such issues? The current answer is: it is very difficult.