aiqbench analysis
GT-HarmBench addresses a critical gap in AI safety by evaluating frontier models on multi-agent coordination failures using game-theoretic structures like the Prisoner's Dilemma. The benchmark reveals that current systems fail to choose socially beneficial actions in 38% of high-stakes scenarios involving military escalation or election manipulation, though game-theoretic interventions can improve outcomes by up to 18%. This standardized testbed provides an essential framework for understanding alignment risks in complex, interactive environments where single-agent metrics fall short.
GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory
arXiv:2602.12316v3 Announce Type: replace-cross Abstract: Frontier AI systems are increasingly capable and deployed in high-stakes multi-agent environments. However, existing AI safety benchmarks largely evaluate single agents, leaving multi-agent risks such as coordination failure and conflict poorly understood. We introduce GT-HarmBench, a benchmark of 1,535 high-stakes scenarios spanning game-theoretic structures such as the Prisoner's Dilemma, Stag Hunt and Chicken. Scenarios are drawn from realistic AI risk contexts in the MIT AI Risk Repository. Across 15 frontier models, agents fail to choose socially beneficial actions in 38% of high-stakes cases, such as military escalation, election manipulation, and medical malpractice. We measure sensitivity to game-theoretic prompt framing and ordering, and analyze reasoning patterns driving failures. We further show that game-theoretic interventions improve socially beneficial outcomes by up to 18%. Our results highlight substantial reliability gaps and provide a broad standardized testbed for studying alignment in multi-agent environments. The benchmark and code are available at https://github.com/causalNLP/gt-harmbench.
Abstract:Frontier AI systems are increasingly capable and deployed in high-stakes multi-agent environments. However, existing AI safety benchmarks largely evaluate single agents, leaving multi-agent risks such as coordination failure and conflict poorly understood. We introduce GT-HarmBench, a benchmark of 1,535 high-stakes scenarios spanning game-theoretic structures such as the Prisoner's Dilemma, Stag Hunt and Chicken. Scenarios are drawn from realistic AI risk contexts in the MIT AI Risk Repository. Across 15 frontier models, agents fail to choose socially beneficial actions in 38% of high-stakes cases, such as military escalation, election manipulation, and medical malpractice. We measure sensitivity to game-theoretic prompt framing and ordering, and analyze reasoning patterns driving failures. We further show that game-theoretic interventions improve socially beneficial outcomes by up to 18%. Our results highlight substantial reliability gaps and provide a broad standardized testbed for studying alignment in multi-agent environments. The benchmark and code are available at this https URL.
Accepted at NeurIPS 2026 Main Conference. Camera-ready will be out soon. This is still the preprint
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY); Computer Science and Game Theory (cs.GT); Multiagent Systems (cs.MA)
arXiv:2602.12316 [cs.AI]
(or arXiv:2602.12316v3 [cs.AI] for this version)
Full text under the publisher's open licence · 193 words
Related coverage
Links #5: 2026/07
LessWrong · Jul 31, 2026
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · Oct 5, 2026
[AINews] High Taste Testers Ep. 1
Latent Space · Oct 17, 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026