Research Summary: GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory

A new research initiative, GT-HarmBench, has developed a benchmark to evaluate artificial intelligence (AI) safety risks in multi-agent, high-stakes environments, addressing a gap in current single-agent focused assessments. The study indicates that frontier AI models frequently fail to select socially beneficial actions in complex scenarios, such as military escalation or election manipulation, revealing significant risks in their deployment within interconnected systems.

Why it matters

The findings highlight a significant vulnerability in current AI systems when operating in multi-agent environments, posing substantial risks to organizational stability and societal welfare. Understanding and mitigating these coordination failures is critical for responsible AI deployment and maintaining trust in increasingly autonomous systems.

Key insights

• Existing AI safety benchmarks primarily focus on single-agent evaluations, overlooking critical multi-agent risks.
• GT-HarmBench is a new benchmark featuring 1,535 high-stakes scenarios based on game-theoretic structures (e.g., Prisoner's Dilemma, Stag Hunt, Chicken) and realistic AI risk contexts.
• Across 15 frontier AI models, agents failed to choose socially beneficial actions in 38% of high-stakes, multi-agent scenarios.
• Specific examples of failure contexts include military escalation, election manipulation, and medical malpractice.
• The research also measures AI agents' sensitivity to how game-theoretic prompts are framed.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2602.12316