GPT-Red is a language model developed by OpenAI to automatically detect vulnerabilities in other AI systems. It tests defenses against attacks like prompt injections and helps improve model security.

What threats does GPT-Red identify?

GPT-Red focuses on prompt injection attacks, where malicious instructions are embedded in text processed by AI. These attacks can lead to data breaches, code corruption, or the generation of harmful content.

How does GPT-Red's training mechanism work?

The model trains through self-play: it attacks other language models while they learn to defend. Over multiple cycles, GPT-Red refines its vulnerability detection skills, and defending models become more resilient.

← All news

Artificial intelligence

OpenAI Develops AI Hacker GPT-Red for Security Testing of Models

July 16, 2026

Photo: MIT Technology Review

Quick answer

OpenAI created GPT-Red, an AI model for automated security testing of language models. The tool identifies novel attack vectors, including prompt injections, and strengthens AI system defenses against malicious actors.

OpenAI researchers have introduced GPT-Red, a tool designed for automated security testing of large language models (LLMs). The development aims to address the growing complexity of AI systems, which are increasingly deployed as autonomous agents interacting with code, web resources, and other models. According to researchers, traditional testing methods struggle to keep pace with the expanding "attack surface" and potential damage from breaches.

GPT-Red primarily targets prompt injection attacks, where malicious actors embed hidden instructions in text processed by AI. Such attacks can result in confidential data leaks, corporate code damage, or the generation of malicious content. The tool has already uncovered previously unknown attack types, enabling proactive security enhancements for models.

GPT-Red was trained using a self-play method: the model interacted with other language models, attempting to attack them while they learned to defend. After several training cycles, GPT-Red significantly improved its vulnerability detection capabilities, and defending models became more resistant to attacks. The creators believe this approach will help prepare for future threats as AI technologies evolve.

Common questions

What is GPT-Red?: GPT-Red is a language model developed by OpenAI to automatically detect vulnerabilities in other AI systems. It tests defenses against attacks like prompt injections and helps improve model security.
What threats does GPT-Red identify?: GPT-Red focuses on prompt injection attacks, where malicious instructions are embedded in text processed by AI. These attacks can lead to data breaches, code corruption, or the generation of harmful content.
How does GPT-Red's training mechanism work?: The model trains through self-play: it attacks other language models while they learn to defend. Over multiple cycles, GPT-Red refines its vulnerability detection skills, and defending models become more resilient.

Dzen feed: /feed/dzen.xml · RSS: /feed.xml