OpenAI Builds GPT-Red — An AI That Finds Vulnerabilities in Its Own Models
OpenAI has developed GPT-Red, an internal system that automates red-teaming of its own LLMs. The model simulates prompt injections and hidden attacks delivered through emails, files, and web pages, then routes them through an adversarial loop: an attacker model and a defender model compete continuously, trained with reinforcement learning and self-play.
In test scenarios, GPT-Red achieved an 84% vulnerability detection rate compared to 13% for human specialists. According to OpenAI, GPT-5.6 Sol’s susceptibility to direct prompt injections dropped 6x compared to models from four months ago — though roughly 3.8% of complex attacks still get through, on par with Claude Opus 4.5. The tool remains internal; OpenAI plans to publish a paper detailing the architecture and training process.