OpenAI uses AI to test its models, outperforming human red teamers
OpenAI's internal GPT-Red model detects 84% of successful attacks in test scenarios, compared to just 13% by human red teamers.
Published 1j1 sourceNotable
Lire en français
≈ 23s
13 %
OpenAI's internal GPT-Red model detects 84% of successful attacks in test scenarios, compare…
The fact
This *self-play training* approach enhances the robustness of future models, such as GPT-5.6 Sol.
The findings are directly used to harden AI systems' security and resilience.
Click the link to read an article on the topic: