OpenAI has developed GPT-Red, an internal AI system designed to test the security of their own AI models. GPT-Red works as an automated "red teaming" tool, looking for security vulnerabilities by sending various prompts to the target AI model and analyzing its responses. Through reinforcement learning self-play techniques, GPT-Red is trained to become more effective at finding weaknesses compared to human security teams. In internal testing, GPT-Red demonstrated a significantly higher success rate in exploiting tested scenarios compared to human teams. Moreover, the use of GPT-Red has also proven capable of reducing the risk of prompt injection attacks and other types of attacks on OpenAI's AI models.
