GT-HarmBench

GT-HarmBench is a game-theoretic AI safety benchmark that evaluates whether large language models can reason strategically in realistic, AI-risk–grounded scenarios.

Each scenario presents two players with a 2×2 payoff matrix embedded in a first-person narrative drawn from real AI risk contexts. Models are evaluated on their ability to:

• identify and play Nash equilibria (individual rationality),
• select actions that maximise utilitarian welfare (sum of payoffs),
• select actions that maximise Rawlsian welfare (min payoff, fairness),
• select actions that maximise Nash social welfare (product of payoffs).

Dataset summary

Total scenarios	2,009
Columns	19
MIT AI Risk Database–sourced	2,009
Game types	6

Game type distribution

Game type	Count
Prisoner's Dilemma	654
Chicken	491
Stag hunt	403
Coordination	252
Bach or Stravinski	170
No conflict	39

Risk level distribution

Scenarios are rated on a 1–10 severity scale.

Risk level	Count
3	8
4	139
5	144
6	673
7	771
8	177
9	89
10	8

Dataset structure

Columns

Column	Description
id	Integer row identifier.
ev_id	MIT AI Risk Database event ID. Empty for synthetic scenarios.
risk_category	Top-level risk category from the MIT AI Risk Database. Empty for synthetic scenarios.
risk_subcategory	Sub-category from the MIT AI Risk Database. Empty for synthetic scenarios.
description	Original risk description from the MIT AI Risk Database. Empty for synthetic scenarios.
formal_game	Canonical game type: one of Prisoner's Dilemma, Chicken, Stag hunt, Coordination, Bach or Stravinski, No conflict.
story_row	First-person narrative given to the row player describing the strategic situation.
story_col	First-person narrative given to the column player.
actions_row	Python-list string of the two action labels available to the row player.
actions_column	Python-list string of the two action labels available to the column player.
1_1_payoff	Payoff vector [row, col] when both players choose action 1.
1_2_payoff	Payoff vector [row, col] when row plays action 1 and column plays action 2.
2_1_payoff	Payoff vector [row, col] when row plays action 2 and column plays action 1.
2_2_payoff	Payoff vector [row, col] when both players choose action 2.
risk_level	Integer 1–10 severity rating of the underlying AI risk.
target_nash_equilibria	Pipe-separated pure Nash equilibria as (row_action, col_action) tuples.
target_utility_maximizing	Pipe-separated action profiles that maximise the sum of payoffs (utilitarian welfare).
target_rawlsian	Pipe-separated action profiles that maximise the minimum payoff (Rawlsian fairness).
target_nash_social_welfare	Pipe-separated action profiles that maximise the product of payoffs (Nash bargaining solution).

Payoff matrix convention

Each row in the dataset encodes a 2×2 game. The four payoff columns (1_1_payoff … 2_2_payoff) each contain a two-element list [row_payoff, col_payoff]. The matrix layout is:

                 col action 1        col action 2
row action 1    1_1_payoff          1_2_payoff
row action 2    2_1_payoff          2_2_payoff

Target columns

Target columns encode optimal action profiles as pipe-separated strings of Python tuples, e.g.:

('go bold', 'go bold')|('play safe', 'play safe')

Multiple profiles appear when ties exist. Each tuple is (row_action, col_action).

Data sources

1. MIT AI Risk Database (ev_id is non-null, 2,009 rows): risk descriptions from the MIT AI Risk Repository were classified as game-theoretic and then contextualized into first-person strategic narratives by an LLM pipeline.

Scenarios were filtered by quality (score ≥ 8/10) and equilibria consistency (score ≥ 8/10) before inclusion. Matching Pennies scenarios (mixed-strategy only, no pure Nash) are excluded.

Usage

from datasets import load_dataset
ds = load_dataset("causalNLP/gt-harmbench")
print(ds[0])

Running the benchmark

# Evaluate a model
uv run python3 -m eval.eval \
    --model-name openai/gpt-4o \
    --dataset data/gt-harmbench-hf.csv \
    --times 1 --temperature 1.0 --experiment-name my-eval

See the GitHub repository for the full evaluation harness.

Citation

@dataset{gt-harmbench,
  title  = {GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory},
  year   = {2026},
  url    = {https://huggingface.co/datasets/causalNLP/gt-harmbench}
}

License

CC BY 4.0