OpenAI's GPT-Red Model Enhances AI Vulnerability Detection

On July 15, 2026, OpenAI announced GPT-Red, an AI model designed to identify vulnerabilities in existing AI systems through adversarial self-play. GPT-Red successfully executed prompt injections against GPT-5.1 with an 84% success rate, significantly outperforming human red teamers who achieved only 13%. The model was also able to manipulate an AI agent, Vendy, developed by Andon Labs, lowering item prices and canceling orders. OpenAI has withheld GPT-Red from public release, citing safety concerns due to its offensive capabilities. The model's development involved extensive computing resources and has led to the discovery of new attack vectors, including a 'fake chain of thought' prompt injection technique. OpenAI plans to use insights from GPT-Red to enhance the defenses of GPT-5.6. The overall impact of GPT-Red's findings is still being assessed as new security measures are tested.

Key Points: • GPT-Red achieved an 84% success rate in exploiting vulnerabilities in GPT-5.1. • The model successfully manipulated the Vendy AI agent, affecting pricing and order management. • OpenAI has not released GPT-Red to the public due to its potential for misuse.