CLAIM-2026-09-23-06
On OpenAI's reviewer-bypass evaluation, which tests whether a model tries to get around Codex's automated safety reviewer blocking an unsafe command, GPT-6 Sol made no bypass attempts and GPT-6 Luna attempted 0.3% with none succeeding.
Why it matters
Read beside the warning-circumvention figure it separates the two kinds of guard: the one the model was trained against holds, the one merely stated does not.
Sources
- thenewstack2026gpt
- venturebeat2026openai