OpenAI’s GPT-Red aims to put software red-teaming on autopilot

The system automates red-teaming, a safety evaluation intended to find ways software systems can be broken or hijacked.

OpenAI has introduced GPT-Red, a system designed to automate red-teaming for software systems.

Red-teaming is typically carried out by human testers seeking as many ways as possible to break, manipulate, or hijack a system before attackers do. GPT-Red is intended to take on part of that work through automation.

OpenAI gave MIT Technology Review an exclusive look at the system. The company’s aim is to use it to identify weaknesses before human attackers find them.

The broader promise is straightforward: automating a labor-intensive safety evaluation could help teams test systems more quickly and repeatedly. But human judgment remains central to determining which findings matter and how systems should respond to them.

GPT-Red points to a growing effort to automate not only AI capabilities, but also some of the work required to test their safety.