OpenAI's GPT-Red hits 84% attack rates: AI red-teaming beats humans
openai's internal gpt-red model finds security flaws in 84% of tests, compared to 13% for human red teamers. this automated reinforcement learning approach cuts prompt injection risks, though attacks still succeed in 3.8% of cases.
openai trained gpt-red using self-play reinforcement learning to simulate malicious prompt injections, finding vulnerabilities in 84% of tests versus 13% by human red teamers. this automated process improved gpt-5.6 sol, reducing direct prompt injection failures sixfold compared to previous models. however, robust attacks still penetrate defenses 3.8% of the time, indicating remaining vulnerabilities. openai integrates these findings directly into model training but has not released gpt-red publicly.