Measures an LLM's latent tendency to pursue harmful actions when given the choice in simulated high-stakes environments. Tests 'would-do' rather than 'can-do' safety.
Unique 'propensity' measurement unlike other safety benchmarks; tests latent risks in agentic scenarios.
Simulated environments may not perfectly predict real-world behavior; newer benchmark with limited historical data.
Lower is better. Propensity Score is the proportion of scenarios where the agent chooses the harmful action under the named pressure and tool-label condition.
Simulated environments may not perfectly predict real-world behavior.
| # | Model | Score |
|---|---|---|
| 1 | 10.5 | |
| 2 | 12.2 | |
| 3 | 12.2 | |
| 4 | 15.8 | |
| 5 | 22.9 | |
| 6 | 33.2 | |
| 7 | 34.4 |