The Multi-Agent Minefield: Can LLMs Cooperate to Avoid Global Catastrophe?
New research maps 2,009 AI risk scenarios to game theory, testing if multiple AI agents can coordinate to avoid disaster.
Deep Dive
Researchers from the GT-HarmBench team published a paper analyzing multi-agent AI safety. They mapped 2,009 high-risk scenarios from the MIT AI Risk Repository onto classic 2x2 games like Prisoner's Dilemma. They found LLMs reached the optimal cooperative outcome 62% of the time, comparable to human baselines (40-60%). The study shows prompting and framing significantly influence whether AI agents achieve cooperative or purely strategic (Nash equilibrium) outcomes.
Why It Matters
As AI agents proliferate, their ability to coordinate, not just act individually, becomes critical for preventing real-world catastrophic failures.