Safety · The Decoder ·

OpenAI is now using AI to attack its own AI, and it's working better than humans ever did

OpenAI is now using AI to attack its own AI, and it's working better than humans ever did

OpenAI says its internal GPT-Red model identified successful attacks in 84% of test scenarios using self-play, compared with 13% for human red-teamers. The findings are being used to harden models, including GPT-5.6 Sol.

Read the full story at The Decoder →