OPENAI
Bot, attack!
OpenAI has created GPT-Red, an internal AI model that searches for security flaws in other GPT systems before they are released.
Human experts already test AI models for weaknesses, but this work takes time and cannot produce enough attacks at scale.
GPT-Red helps automate the process by sending harmful prompts, studying the response and trying again until it finds a weakness.
It mainly tests prompt injection, where hidden instructions inside emails, websites, files or tool results try to make an AI perform harmful actions.
OpenAI used GPT-Red’s attacks to train GPT-5.6 Sol.
The company says the model had six times fewer failures on its hardest direct prompt-injection test than its strongest production model from four months earlier.
In brief:
GPT-Red automatically searches for AI security flaws.
Its attacks helped make GPT-5.6 Sol harder to manipulate.
It will support human red-teamers rather than replace them.
Security hired a menace
In another test, GPT-Red successfully attacked 84% of new scenarios, compared with 13% for human red-teamers.
OpenAI plans to keep GPT-Red private and use it alongside human testing, external researchers and other safety systems.