OpenAI internal circumvention - reported rate
Safety
Attempted circumvention of auto-review denials in an internal OpenAI evaluation. Task cohort and grading are unreported; diagnostic, lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 0.0 |