OpenAI has disclosed details of an internal automated red-teaming system called GPT-Red that it used to identify and fix prompt injection vulnerabilities in its GPT-5.6 language model. The system represents a significant step in using AI to improve the security of other AI systems.
GPT-Red is a specialized AI model designed specifically to probe other models for security weaknesses. OpenAI said it deployed GPT-Red against earlier versions of its models and found them highly vulnerable to prompt injection attacks, where malicious instructions are embedded within seemingly benign inputs.
“GPT-Red is a strong red-teamer, and our previous models are highly vulnerable to its prompt injection attacks,” OpenAI stated. The company used the findings from GPT-Red to adversarially train GPT-5.6, making it significantly more resistant to such attacks.
Prompt injection has emerged as one of the most pressing security concerns in the AI industry. These attacks can trick AI models into ignoring their safety guidelines, generating harmful content, or revealing sensitive information. The technique has been demonstrated against most major language models, and defending against it has proven challenging.
The automated approach represents a departure from traditional red-teaming, which relies on human security experts manually testing systems. While human red teams remain valuable for identifying complex or novel attack vectors, automated systems like GPT-Red can conduct testing at vastly greater scale and speed.
OpenAI has been increasingly focused on AI safety as its models become more capable. The company has deployed multiple layers of defense, including system-level safeguards, training-time mitigations, and post-deployment monitoring. GPT-Red adds an additional automated testing layer that can continuously evaluate new model versions as they are developed.
The approach mirrors similar efforts at other AI labs. Anthropic and Google DeepMind have also published research on automated safety testing, and the industry is moving toward shared standards for evaluating model robustness.
This article was adapted from Decrypt. Read the original here.
