OpenAI unveils GPT‑Red, the formidable AI that hacks ChatGPT, but it’s for its own good

OpenAI unveils a new AI model. Named GPT‑Red, the artificial intelligence model is solely designed to find flaws in the startup’s other AI systems before cybercriminals do.To uncover vulnerabilities, AI will simulate real cyberattacks against the models powering ChatGPT. These simulations, known as “red teaming” in the cybersecurity world, are intended to isolate weak points in a computer system.

Introducing GPT-Red

An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scale, helping us build stronger defenses before wider deployment.https://t.co/GxnmxxcpSk

— OpenAI (@OpenAI) July 15, 2026

Until now, this work was primarily carried out by researchers. With its GPT‑Red model, the startup decided tofully automate the process, thereby accelerating the detection of flaws. OpenAI says that GPT‑Red was trained with computing power comparable to that of its largest AI models. This is an “unprecedented” level of training for a tool dedicated exclusively to security. Sam Altman’s company hopes that GPT-Red will eventually become more autonomous and learn to detect its own vulnerabilities, before they are exploited by third parties.

Also read: ChatGPT locks down against cyberattacks – this new mode will better protect your data; how to activate it?

Prompt injection attacks in the spotlight

The AI model is primarily designed to prevent prompt injection attacks.This attack technique relies on inserting malicious commands directly into a request submitted to the AI. The system then interprets these instructions as legitimate and, if they are sufficiently precise, complies with them without resistance. For hackers, the goal is to push the AI to ignore its own guardrails in order to carry out harmful actions. The fundamental problem remains that large language models still fail to reliably distinguish between the content they analyze and the orders they are given.

This risk has grown considerably with the proliferation of AI agents capable of browsing the web, reading files, or using third-party tools autonomously. Agents multiply the potential entry points. This is why OpenAI decided to act by developing GPT‑Red.

Also read: The ChatGPT update we were all waiting for – the AI now speaks like a human, it’s impressive

An AI that trains by fighting other AIs

To train GPT-Red, OpenAI pitted its AI against other AI models designed for defense. Each time GPT‑Red manages to breach the defenses of one of them, it is rewarded, while the defender learns to strengthen itself against this type of attack. This mechanism drives all models to improve and evolve simultaneously.In effect, GPT‑Red must constantly devise more sophisticated strategies to continue breaking through adversarial defenses, driving it to become increasingly effective.

To test GPT‑Red under real-world conditions, OpenAI pitted it against Vendy, an AI agent responsible for a vending machine located on its premises. With little effort, GPT‑Red managed to manipulate the prices of products sold by the vending machine. It was also able to cancel an order placed by another customer, or claim an item for itself for only $0.50.

The discovery of GPT-Red

During this training process, OpenAI discovered an entirely new category of attacks, dubbed “fake chain-of-thought” (false chain reasoning). This technique consists of convincing the model that it has already checked and validated false information throughout its own internal reasoning, so that it subsequently repeats it without questioning it. In plain terms, the attack amounts to persuading a model that 1+1 equals 3, simply by making it believe that it has itself already performed this calculation and arrived at this result.

As OpenAI explains in its press release, GPT-Red also proved markedly more effective than human researchers. The AI succeeded in breaking through the defenses in 84% of the scenarios tested, compared with only 13% for human experts. The startup emphasizes that using GPT-Red helped improve ChatGPT’s resilience.While more than 90% of Red’s attacks worked against GPT‑5, this rate fell to less than 23% with the GPT‑5.6 update, demonstrating the progress made in security.

GPT-Red’s blind spots

However, the king of AI acknowledges that the tool still has blind spots. Despite its elite training, the model remains ineffective against long, progressive attacks, as well as offensive techniques hidden in images. Human cybersecurity experts therefore remain indispensable. To prevent abuse, OpenAI has chosen never to make GPT‑Red available to the public. The AI is strictly reserved for internal use.

GPT-Red embodies OpenAI’s cybersecurity strategy. The American startup aims to leverage today’s artificial intelligence to build future generations of safer, more resilient AI models. The company describes this mechanism as a “virtuous cycle”.

👉🏻 Follow the latest tech news in real time: add 01net to your feeds on Google, and subscribe to our WhatsApp channel.