A new Stanford research project called 'Agent Island' uses Survivor-style elimination games to benchmark AI behavior, addressing the problem of saturated and contaminated traditional evaluations. The study, published by Connacher Murphy of the Stanford Digital Economy Lab, involved 49 AI models competing in 999 simulated games. Models negotiate alliances, manipulate votes, and eliminate rivals, revealing dynamics like same-provider favoritism. OpenAI's GPT-5.5 ranked first with a skill score of 5.64, while Anthropic's Claude Opus models also performed well. The project highlights the need for dynamic benchmarks as AI agents gain autonomy and may pursue conflicting goals. The study notes potential dual-use risks, as logs could improve AI persuasion and coordination strategies.

Key facts

• Agent Island tests AI models with Survivor-style multiplayer elimination games.
• GPT-5.5 ranked first among 49 AI models with a skill score of 5.64.
• AI models showed a 8.3% bias toward finalists from the same provider.
• The benchmark aims to overcome saturation and data contamination in traditional tests.

KeyAudit data perspective