Safety Behavior Prediction on Assistant Turns
88.7Production RateDeployment Simulation
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Sexual content2026.07 | 88.7 | 1.2 | 100.3 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Harassment2026.07 | 70.4 | 1.3 | 34.2 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Sexual content2026.07 | 69.4 | 1.3 | 58.4 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Harassment2026.07 | 56 | 1.1 | 37.6 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Sexual content2026.07 | 53 | 1.7 | 56.2 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Hate2026.07 | 38 | 2.1 | 37.9 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Violent wrongdoing2026.07 | 35.2 | 1.8 | 26.1 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Harassment2026.07 | 32 | 2.2 | 27.5 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Non-violent wrongdoing2026.07 | 31 | 1.3 | 27.9 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Violent wrongdoing2026.07 | 24.9 | 1.6 | 26.2 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Non-violent wrongdoing2026.07 | 23.6 | 1.8 | 30.4 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Violent wrongdoing2026.07 | 19.7 | 1.3 | 14.8 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Hate2026.07 | 17.8 | 2.2 | 14.3 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Non-violent wrongdoing2026.07 | 13.2 | 1.2 | 19.7 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Hate2026.07 | 11.4 | 3.3 | 12.2 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Violence2026.07 | 11.4 | 1.2 | 9.5 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Violence2026.07 | 11.3 | 1 | 10 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Sexual minors2026.07 | 9.4 | 1.3 | 12.6 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Violence2026.07 | 7.7 | 1.5 | 5.8 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Sexual minors2026.07 | 7.2 | 1.7 | 7.3 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Self-harm2026.07 | 5.6 | 1 | 5.8 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Self-harm2026.07 | 5.5 | 1.1 | 7.3 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Self-harm2026.07 | 5 | 1.4 | 6.2 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Sexual minors2026.07 | 2.4 | 3.9 | 3 | — | — | |
| Deployment SimulationDeployment=GPT-5.1 → GPT-5.2, Category=Extremism2026.07 | 2 | 2.5 | 1.9 | — | — | |
| Deployment SimulationDeployment=GPT-5.2 → GPT-5.4, Category=Extremism2026.07 | 0.8 | 5.6 | 0.8 | — | — | |
| Deployment SimulationDeployment=GPT-5 → GPT-5.1, Category=Extremism2026.07 | 0.1 | 17.9 | 0.7 | — | — |