Flow - 4x faster than typing
Mini toolsGPT Red
By OpenAI
Model family: GPT
GPT-Red is OpenAI's most advanced automated safety red-teaming model, built to scale vulnerability discovery beyond what human red-teamers can achieve. It is trained through self-play reinforcement learning: GPT-Red and a population of diverse defender models are trained simultaneously across a broad set of prompt-injection scenarios, with GPT-Red rewarded for eliciting failures and defenders rewarded for resisting them.
GPT-Red operates like a human red-teamer, sending a prompt, observing the model's response, and iterating, but at the compute scale of some of OpenAI's largest post-training runs. Its attacks are then used to adversarially train OpenAI's production models; GPT-5.6 Sol, trained against GPT-Red, achieved 6x fewer failures on OpenAI's hardest direct prompt injection benchmark compared to the prior production model, and fails on only 0.05% of GPT-Red's direct injection attempts.
In evaluations, GPT-Red found successful attacks on 84% of a novel indirect prompt injection benchmark versus 13% for human red-teamers, and was used to break live agentic systems in case studies, including an AI-powered vending machine built by Andon Labs and a Codex CLI coding agent, exposing vulnerabilities that were then disclosed and patched.
GPT-Red is deliberately kept separate from deployed models and is never released to users, so its attack capabilities stay out of the hands of adversarial actors while still being used to make OpenAI's shipped models more robust.
GPT-Red operates like a human red-teamer, sending a prompt, observing the model's response, and iterating, but at the compute scale of some of OpenAI's largest post-training runs. Its attacks are then used to adversarially train OpenAI's production models; GPT-5.6 Sol, trained against GPT-Red, achieved 6x fewer failures on OpenAI's hardest direct prompt injection benchmark compared to the prior production model, and fails on only 0.05% of GPT-Red's direct injection attempts.
In evaluations, GPT-Red found successful attacks on 84% of a novel indirect prompt injection benchmark versus 13% for human red-teamers, and was used to break live agentic systems in case studies, including an AI-powered vending machine built by Andon Labs and a Codex CLI coding agent, exposing vulnerabilities that were then disclosed and patched.
GPT-Red is deliberately kept separate from deployed models and is never released to users, so its attack capabilities stay out of the hands of adversarial actors while still being used to make OpenAI's shipped models more robust.
New
Text
Overview
GPT-Red is OpenAI's internal automated red-teaming model, trained via self-play reinforcement learning to discover prompt injection and other safety vulnerabilities at scale. It generates adversarial attacks used to adversarially train production models such as GPT-5.6, making them more robust. GPT-Red is never deployed to users; it is kept internal-only.
About OpenAI
OpenAI is a technology company that specializes in artificial intelligence research and innovation.
Industry:
Artificial Intelligence
Company Size:
4500
Location:
San Francisco, California, US
Website:
openai.com
Last updated: July 16, 2026
