U.
We know models cheat.
OpenAI is discussing earlier AI safety evaluations with groups including METR and Redwood Research as it expands independent model testing.
Eight of the ten most-used models on OpenRouter in August ran on open weights.
CLOSEDQUORUM uses votes from up to four AI models to choose theft, injection, or persistence, but its public build is nonfunctional.