OpenAI Uses GPT-Red to Cut GPT-5.6 Sol Prompt-Injection Failures 6x
Updated
Updated · OpenAI · Jul 15
OpenAI Uses GPT-Red to Cut GPT-5.6 Sol Prompt-Injection Failures 6x
3 articles · Updated · OpenAI · Jul 15
Summary
GPT-5.6 Sol posted 6x fewer failures on OpenAI’s hardest direct prompt-injection benchmark than its best production model from four months earlier after being adversarially trained with GPT-Red.
GPT-Red is an internal automated red-teaming model trained via self-play reinforcement learning to iteratively attack defender models, addressing a scaling bottleneck in human-led safety testing.
84% of novel test scenarios were successfully attacked by GPT-Red versus 13% for human red-teamers, and OpenAI said it can break nearly all models it was pitted against up to GPT-5.5.
0.05% of GPT-Red direct prompt injections now cause GPT-5.6 Sol to fail, while earlier “Fake Chain-of-Thought” attacks that topped 95% success on GPT-5.1 are below 10% on the new model.
OpenAI said it keeps GPT-Red separate from deployed products, is testing safeguards for vulnerabilities it exposed, and plans a detailed preprint later this week.
Has OpenAI's AI 'super-hacker' just ignited an unavoidable AI vs. AI arms race?
A creator of OpenAI's hacker AI also designed its target. Does this 'insider job' overstate its real-world effectiveness?
Automated Red-Teaming and GPT-5.6: Raising the Bar for AI Security, Regulation, and Human Oversight
Overview
Automated red-teaming is transforming AI security by making it more scalable and efficient. As advanced models like OpenAI’s GPT-5.6 Sol become more capable, traditional human-led red teaming—valued for its creativity and ability to spot new threats—has become too resource-intensive for many organizations. OpenAI has responded by developing automated red-teaming methods, sharing research to strengthen industry practices and improve safety. By combining human ingenuity with automation, organizations can better identify vulnerabilities in rapidly evolving AI systems, ensuring that security keeps pace with technological progress while making robust testing accessible to more teams.