OpenAI showcases GPT-Red: an internal AI model that attacks other models with prompt injections. The company used the model to train GPT-5.6 Sol.
OpenAI has developed GPT-Red: an internal security model designed to test other GPT models. With this model, OpenAI aims to make its own AI systems more resilient to prompt injection. AI models process a lot of external data via browsers, linked applications, local files, and tools to perform tasks, but this also makes it possible to place hidden instructions in emails, web pages, or code repositories.
Such a prompt injection might, for example, try to convince a model to send sensitive data to an external server. Human red teams track down such errors, but according to OpenAI, they do not generate enough attacks to train security at scale.
Attacker and defender train together
GPT-Red works like a human attacker: the model sends a prompt, observes the response of a target model, and then refines its attack. OpenAI trained the model with computing power on par with some of its largest models, exclusively for security research.
The training relies on self-play with multiple defensive language models. GPT-Red receives a reward when it causes a valid error, while the defenders are rewarded if they repel the attack and perform their original task. As a result, GPT-Red must constantly find new attack techniques.
OpenAI built scenarios for this in which the attacker controls parts of a local file, web banner, email, or tool output. The company keeps GPT-Red separate from models it offers externally, because the model deliberately learns harmful capabilities.
Strong results and real vulnerabilities
In a simulated test environment for indirect prompt injections, GPT-Red successfully attacked GPT-5.1 in 84 percent of the scenarios. Human red-teamers achieved 13 percent in the same test. According to OpenAI, the scenarios differed from the environments on which GPT-Red was trained.
The model also tested an AI-driven vending machine in an OpenAI office. GPT-Red was able to lower the price of an expensive product to $0.50, offer a new product worth more than $100 for that price, and cancel another customer’s order. OpenAI reported the vulnerabilities and is testing new safeguards.
In ten separate data exfiltration scenarios, GPT-Red also proved more effective and token-efficient than a GPT-5.5 base model when attacking a Codex CLI agent with GPT-5.4 mini.
GPT-5.6 must better withstand attacks
OpenAI has been using predecessors of GPT-Red since GPT-5.3 during the training of successive production models. An early version discovered, among other things, attacks using a forged chain of reasoning. These succeeded with GPT-5.1 in more than 95 percent of cases, but remain below ten percent with GPT-5.6 Sol. According to OpenAI, GPT-5.6 Sol fails in only 0.05 percent of direct prompt injections from GPT-Red, without refusing legitimate requests.
OpenAI intends to use more computing power and data to make future versions of GPT-Red stronger. Automated attacks do not replace human red teams, according to OpenAI, but must work alongside layered security and real-time monitoring to prevent vulnerabilities from appearing only after a broad rollout. Whether OpenAI will ever release this model remains to be seen, but with GPT-5.5-Cyber, it offers a model that companies can use to detect vulnerabilities.
